dsh-continual-harness: A Continual Self-Evolution Plugin for DeepSeek Harness

dsh-continual-harness is a continual self-refinement plugin for DeepSeek Harness (DSH plugin). It gives the agent a closed loop of persistent memory, periodic review-and-refine, cross-session shared knowledge, and automatic rollback on failure, implemented through DSH's plugin mechanisms (session events, agent-scoped events, pre-step waterfall, tools service). Inspired by Prime Intellect's open-source prime-agent, it addresses the problems of agent forgetting, lack of knowledge sharing, and uncontrollable optimization in long-running tasks. Core capabilities include state projection, automatic refinement, manual refinement tool, memory lifecycle management, ranked injection, session wrap-up, and invariant guard.

What core features does the dsh-continual-harness DSH plugin support?

This plugin is a single npm package (dsh-continual-harness) that takes effect through multiple extension points once mounted: state projection (injects harness context each step via pre-step waterfall listener, with incremental injection when content digest changes), review and automatic refinement (session/event listener on turn interval or compaction end, runs LLM review → plan → apply automatically), manual refinement tool (registers the harness_refine tool, directly callable by the LLM, supports rollback), memory lifecycle (manual archive/unarchive/pin through refinement metadata; archived entries are hidden from injection and skill materialization), ranked injection (queries the latest effective direct-user message up to 400 chars, ranks title matches above content matches, then applies freshness/id tie-breaks and a per-kind cap), session wrap-up (optional harness_wrapup tool gives mechanical keep/promote/archive advice; promotion is copy-only and conflicts return a deterministic error), in-session review trajectory (rebuilt from session logs with tail-biased truncation), and invariant guard (harness/refinement event validation plus batched failure reporting).

How to install dsh-continual-harness?

Use the DSH plugin management command:

dsh plugin --profile web add github:jasen215/dsh-continual-harness

Once installed, the plugin automatically registers with DSH's plugin mechanisms and takes effect without additional configuration.

What is the data storage and architecture of dsh-continual-harness?

The plugin stores data in the shared ESP experience root (default ~/.dsh/harness/), including harness_state.json (cross-session global state), refinements.jsonl (append-only global refinement history), reviews.jsonl (cross-batch gate/audit history), and continual-harness.log (implementation log, JSONL, 0600 permissions). Architecturally, the src/ directory contains modules: domain.ts (event declaration merging), types.ts (type definitions), storage.ts (disk read/write with atomic writes, corruption degradation, local/global merge), refine.ts (validation, application, rollback with baseline conflict detection, version increments, growth limits), skills.ts (SKILL.md rendering and file reconciliation, generated skills are real dsh skills), render.ts (model-facing overview/summary/history rendering), usage.ts (injection telemetry), wrapup.ts (deterministic session wrap-up suggestions), planner.ts (LLM planning prompts and JSON parsing), store.ts (HarnessStore combining storage and event publishing), complete.ts (completion via ctx.get('llm')), tool.ts (harness_refine tool), projection.ts (pre-step projection with digest dedup), driver.ts (automatic refinement driver with turn-interval gate, compaction gate, cooldown, re-entry guard), invariant.ts (runtime invariant plugin), and index.ts (plugin entry and Config). Tests cover 17 test files and 163 cases.

Which DeepSeek Harness versions and platforms does dsh-continual-harness support?

Maintained by jasen215, this plugin is licensed under MIT and written in TypeScript. It depends on official DSH packages such as @deepseek-ai/dsh-agent, @deepseek-ai/dsh-session, and @deepseek-ai/dsh-tools, ensuring compatibility with DeepSeek Harness. The plugin was created on 2026-08-17 and last updated on 2026-08-19, with 4 stars and 0 forks. It is suitable for agents that require long-term self-evolution, such as continuous coding assistants and knowledge management agents.

View more DSH plugins at dsh-plugin.org