dsh-self-evolve
Self-evolving plugin for DeepSeek Harness: observes agent behavior, proposes improvements to its own skills/policies via LLM, validates them in a sandboxed trial agent, and applies versioned mutations with rollback.
安装
$
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:lispking/dsh-auto-evolve说明文档
阅读完整 README ↗Custom configuration
The bundle applies a default config; to change it, override the row in your own profile patch (applied after every bundle layer):
### Configuration
| Field | Default | Meaning |
|---|---|---|
| `mode` | `observe` | Evolution mode (see above). |
| `observation.toolFailureThreshold` | `3` | Tool-failure burst count that triggers a cycle. |
| `observation.repeatThreshold` | `3` | Identical-call count treated as a no-progress loop. |
| `observation.requestErrorThreshold` | `3` | LLM request-error count that triggers a cycle. |
| `observation.windowMs` | `300000` | Rolling window (ms) over which signal counts aggregate. |
| `proposal.maxProposalsPerTrigger` | `1` | Max mutations per proposal. |
| `proposal.maxEpisodesPerProposal` | `5` | Max observations rendered into the proposal prompt. |
| `proposal.maxPromptChars` | `24000` | Prompt size cap (bounds cost). |
| `proposal.maxTokens` | `2000` | Max output tokens for one proposal call. |
| `validation.maxTrialMs` / `maxToolCalls` | `30000` / `20` | Trial wall-clock and tool-call caps. |
| `validation.maxTrialSteps` / `maxTrialTokens` | `12` / `8000` | Trial model-step and per-request token caps. |