dsh-evolver
面向 DeepSeek Harness 的可审计、由验证器门控的自我演进。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:cofy-x/dsh-evolver#62442c658ba3a46b2d2dcfef027b7204f46475ebEvolver 以 dsh-evolver 发布，当前版本为 0.1.0-alpha.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Auditable, verifier-gated strategy evolution for DeepSeek Harness.
DSH Evolver turns bounded execution failures into reviewable strategy proposals. Verification, human acceptance and explicit promotion must all happen before guidance reaches a later Session. It does not autonomously edit DSH, plugins or user code. The product is experimental: its fixed diagnostic template and structural verifier do not establish improved task success.
Install the published alpha into a DSH profile, then restart that profile:
dsh plugin --profile web add dsh-evolver@alpha
dsh --profile web --dump-config
The archive includes built JavaScript and declarations. Replace alpha with a published version for an exact candidate. package.json declares the development baseline and host compatibility; the registry channel may still point to an earlier release. Git installation remains available as dsh plugin --profile web add github:cofy-x/dsh-evolver and requires the development toolchain.
After a model-requested tool fails, use a command-capable DSH client:
/evolve list
/evolve patterns
/evolve pattern <pattern-id>
/evolve show <proposal-id>
/evolve accept <proposal-id>
/evolve promote <proposal-id>
/evolve evaluate <proposal-id>
Acceptance and promotion are separate human decisions. Reject with /evolve reject <proposal-id> <reason>; withdraw active guidance with /evolve rollback <proposal-id> <reason>. Promotion and rollback affect later Sessions without erasing history.
The offline plugin exposes ctx.evolver, groups redacted failures into exact-signature patterns and manages proposal generations. Only promoted text enters model context through the public, logged agent.inject() channel. Evaluation is an exposure-scoped failure-rate comparison that informs human decisions, not automatic rollback or proof of causality.
The audit stores bounded diagnostic facts and canonical Session references, never complete transcripts, reasoning, tool arguments, credentials or complete tool output. Redaction is defensive, not a guarantee that arbitrary text is nonsensitive; protect the owner-only audit file. Use one active runtime per data directory.
Configuration is defined by the loader schema. The profile patch adds an optional plugin row to Web or Headless profiles. Evolver owns neither model credentials nor tool permissions and cannot execute proposal text or rewrite verifier, evaluator, approval or safety policy.
Use Node.js 24 or newer and pnpm 11:
git clone https://github.com/cofy-x/dsh-evolver.git
cd dsh-evolver
pnpm install --frozen-lockfile
pnpm run build
Use pnpm run prepare:harness to build the pinned released Harness source for integration tests. Follow AGENTS.md and the verification guide. The documentation map routes architecture, release and experiment work. Architecture defines future priorities and extension prerequisites: establish useful strategy evidence before adding providers, automation or storage complexity.
This independent, MIT-licensed DSH-native implementation is inspired in part by EvoMap/evolver and experience-driven agent-evolution research. It is not affiliated with or endorsed by EvoMap; no EvoMap source code, prompts or private formats are implementation inputs, and no protocol compatibility is claimed. See LICENSE.