dsh-continual-evolve DSH plugin

dsh-continual-evolve enables DeepSeek Harness agents to persistently store and refine experience across sessions. It provides code-enforced safety for versioned entries, deterministic rollback, and an automatic review gate with human approval for global edits. The plugin also includes a memory layer with ranked injection, trajectory citations, and archive functionality.

developer-tools skills-workflows

How to install the dsh-continual-evolve DSH plugin

dsh plugin --profile web add dsh-continual-evolve

Copying does not run this command. Review the repository and version before installing the dsh-continual-evolve DSH plugin.

dsh-continual-evolve DSH plugin data source

dsh-continual-evolve DSH plugin snapshot date: Aug 16, 2026

discovered

What the dsh-continual-evolve DSH plugin can do

Versioned entries for prompt notes, memories, skills, and subagent specs with provenance tracking
Deterministic rollback via inverse-op edits (no LLM re-guessing)
Code-enforced safety: schema validation, atomic writes, corrupt-file degrade, optimistic concurrency, immutable base system prompt
Local (session) and global (cross-session) scopes with merge semantics
Memory layer: ranked injection (BM25+recency), trajectory citations, archive/unarchive
Auto-review gate with turn-interval/compaction checkpoints and human approval for global edits
Benchmark-driven validation loop with code-owned scoring and non-regressive acceptance
Skill materialization and hot-mounting as live cordis plugins
Plugin-owned file logging (JSONL exporter with rotation)

Where the dsh-continual-evolve DSH plugin fits

Persist useful agent memories and skills across sessions so they are not lost at next turn
Audit and rollback any refinement with a full event trail
Automatically review and approve/reject self-evolution proposals with a configurable gate
Hot-mount a skill entry as a live plugin exposing a tool to the agent
Export/import store for backup or migration between DSH installations

Who the dsh-continual-evolve DSH plugin is for

Developers building long-running DeepSeek Harness agents that need persistent memory and skill refinement
Researchers or engineers who want a production-grade, code-enforced self-evolution loop for their agents

dsh-continual-evolve DSH plugin limitations

Requires DeepSeek Harness environment and Node.js ^22.19.0 || >=24.0.0 (matches DSH)
Memory injection capped at 6 entries per kind (with ranked selection)
Skill standard in the loop depends on optional skill-creator/skill-audit skills being installed; otherwise falls back to a builtin distilled guide (~1KB)
The auto-review gate uses raw ctx.llm calls and does not load agent skills (skill tool not available in planner/gate context)
Plugin is in long-term maintenance; no new phases planned but usage-driven enhancements continue

dsh-continual-evolve DSH plugin: from the repository README

Quoted from the ZK-Andy/dsh-continual-evolve README, the upstream source of the dsh-continual-evolve DSH plugin. Copyright remains with the original authors.

Continual self-evolution for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness): a versioned, auditable, rollback-safe layer of harness state — prompt notes, memories, skills, and subagent specs — refined from session trajectories. > **Status: all phases complete; in long-term maintenance.** Phases 1–3 > shipped the full evolution loop: the pure-core engine, model tools and > the `/evolve` command, the 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 the benchmark-driven validation loop (code-owned > scoring, non-regressive acceptance, rubric ACL). Since then the plugin > keeps growing with usage-driven enhancements — the memory layer (ranked > injection, trajectory citations, archive), per-installation rubric keys, > and plugin-owned file logging. See the Roadmap for the full shipped and > candidate lists. ## Background This project started as a research question: *can a harness improve itself, and what would a production-grade version look like?* Three lines of evidence shaped the answer: - **penguin-harness** d

Read the full READMERepository license: MIT

dsh-continual-evolve DSH plugin questions

How do I install dsh-continual-evolve?

Run `dsh plugin --profile web add dsh-continual-evolve` to install from npm. Alternatively, install from source with `dsh plugin --profile web add github:ZK-Andy/dsh-continual-evolve`. Replace `web` with your actual profile name (e.g., headless). After installation, restart DSH for the plugin to activate.

Can I roll back a refinement?

Yes. Use `/evolve rollback <id>` where `<id>` is the refinement ID shown in `/evolve history`. The rollback is deterministic: inverse edits are generated from the applied results, not by asking the LLM to guess. The rollback is a new event in the audit trail, so you can always revert again.

Does this plugin require an external memory service?

No. It is entirely self-contained with pure functions. The memory layer uses ranked injection (BM25+recency) and trajectory citations, all computed locally. There is no dependency on external services like Mem0, Letta, or Zep.

How do I see what the plugin has evolved?

Use `/evolve list` to list all entries, or `/evolve list global` for global entries. `/evolve history` shows the applied refinements with IDs. You can also use `/evolve log` to view the plugin's own file log (JSONL).

What happens if the store file gets corrupted?

The plugin implements corrupt-file degrade: if the store file is corrupted, it will degrade gracefully rather than crash. The state module reads the store with integrity checks and can fall back to a safe state. If you have a backup, you can restore it via `/evolve import <path>`.