huchunlinnk/deepseek-desk-rsi0

dsh-desk-rsi

Recursive Self-Improvement (RSI) engine for DeepSeek Harness: a bounded perceive → integrate → verify → repair → propose loop that keeps a downstream desktop app in sync with the fast-moving upstream harness. AI for AI — DSH maintains DSH.

包名
dsh-desk-rsi
版本
0.1.0
许可证
MIT
最近更新
2026年8月18日

安装

$npx -p @deepseek-ai/dsh dsh plugin --profile web add github:huchunlinnk/deepseek-desk-rsi

deepseek-desk-rsi

License: MIT

AI for AI: a DeepSeek Harness (DSH) plugin that makes a downstream app keep itself in sync with the fast-moving upstream harness — using DSH itself as the worker. DSH maintains DSH.

DeepSeek Harness is in 0.1.x-rc and its authors promise breaking changes. A desktop shell that pins one rc version rots within days. Instead of a human chasing upstream, this bundle gives the agent a bounded, self-correcting loop over five primitives.

The loop

perceive  →  integrate  →  verify  →  parity  →  repair (bounded)  →  propose
StageWhoHow
perceivetool rsi_perceivefetch upstream, diff commit log + stat since last sync
integrateagent (existing tools)apply relevant harness changes to the app's glue with bash / edit / subagent / workflow
verifytool rsi_verifythe hard fitness function: cargo build --locked && cargo test + smoke, ok only on exit 0
paritytool rsi_paritythe identical-surface gate: HTTP endpoints answer and every required plugin in parity.json is still present
repairagent + rsi_checkpoint / rsi_rollbackcheckpoint before edits (Memento), roll back and retry after a failed verify/parity
proposetool rsi_proposecommit, push a branch, open a PR via gh — the PR is the human gate; the loop never merges

The design constraint that makes RSI tractable here: verify + parity have hard, automatable reward signals ("does it build and pass", "is every original capability still present"), unlike vague "get smarter" objectives.

Design principles

  • First principles — the bundle reuses the harness (@deepseek-ai/dsh-tools, @deepseek-ai/cordis are peer deps provided by DSH); it never reimplements bash, git, or subagents.
  • Occam's razor — plain ESM JavaScript, zero build step, zero prepare script; a git install works out of the box.
  • Design patterns — the loop is a Pipeline; checkpoints are Memento; the PR is a Command gate; the repair loop is bounded recursion.

Install

dsh plugin --profile rsi add github:huchunlinnk/deepseek-desk-rsi#main
dsh --profile rsi "Run the daily upstream sync: perceive, integrate, verify, repair if needed, then propose a PR."

The profile/ directory is a starting headless profile; copy it into $DSH_HOME/profiles/rsi and set the git remote / verify command for your repo.

Tools

  • rsi_perceive — diff upstream since a base ref.
  • rsi_checkpoint — commit a labeled snapshot (Memento capture).
  • rsi_rollbackgit reset --hard back to a checkpoint (Memento restore).
  • rsi_verify — run the fitness command; the build/test truth signal.
  • rsi_parity — check HTTP endpoints + required plugins against parity.json; the 1:1 feature-parity gate.
  • rsi_propose — open a PR via gh; the human review gate.

Parity contract

parity.json lists every plugin in the official dsh-base + dsh-web-app bundles (128 plugins) plus the host endpoint. rsi_parity fails if any name goes missing after a sync — the machine-readable "identical surface" check. Regenerate it from the official bundles whenever upstream adds/renames a plugin:

node scripts/gen-parity.mjs /path/to/deepseek-harness

Verify locally

mkdir -p node_modules/@deepseek-ai
ln -s /path/to/deepseek-harness/vendor/cordis node_modules/@deepseek-ai/cordis
ln -s /path/to/deepseek-harness/packages/core/tools node_modules/@deepseek-ai/dsh-tools
node scripts/smoke.js      # load + register all tools and the loop prompt
node scripts/e2e-loop.mjs  # full perceive → checkpoint → rollback → verify → parity on real git

License

MIT