lispking/dsh-auto-evolve

dsh-auto-evolve is a self-evolving plugin for DeepSeek Harness: it owns a genome of evolvable assets (skills, tool-result post-processors, prompt sections, guard policies), collects runtime signals such as tool failures and repeated calls, drafts a mutation through ctx.llm, replays the failing episode in a sandboxed trial agent and applies a change only when the trial beats the baseline - with a versioned ledger and automatic rollback.

dsh-auto-evolve is a self-evolving plugin for DeepSeek Harness. Instead of shipping a static skill set, it owns a small genome of evolvable assets - skills, tool-result post-processors, prompt sections and guard policies. It observes runtime signals (tool failures, no-progress repeats, request errors) on tools/result and agent/request-error, and when a threshold crosses, a proposal cycle snapshots the genome and calls ctx.llm.stream() with a strict prompt, validating the output against a closed mutation vocabulary (add/patch/retire over those four kinds) so the model never invents asset kinds or operators. A sandboxed trial agent replays the failing episode twice - baseline and with the candidate - and compares completion, tool failures and tool-call cost, applying the change only when the trial wins. Every applied mutation is versioned in a durable ledger with its previous content captured and a disposer kept, so a regression rolls back to the exact prior state; prompt-section candidates are recorded but not auto-applied.

Agent Capabilities ★ 3
View on GitHub ↗

Install

dsh plugin --profile web add dsh-auto-evolve

Copied verbatim from the README's Installation block; a TUI-profile line (dsh plugin --profile tui add dsh-auto-evolve) and a local source-build form are also documented.

Compatibility

A DSH bundle: dsh plugin add initializes the profile if needed, installs the package and adds dsh-auto-evolve to the profile's bundle stack automatically, after which the bundled cordis.patch.yml registers the plugin row with its defaults. Observe-only is the default mode; a budget gate (maxCostPerCycle / dailyBudget) bounds the proposal call and every trial replay in auto-apply.

Details

Alternatives

030611/dsh-telemetry-redactor · 173787247/dsh-wsl-env · acnlabs/dsh-plugin-acn

More plugins in Agent Capabilities

Browse more in Agent Capabilities

Guides for Agent Capabilities plugins