DSH HUB
HomePlugin StorePlugin PacksCommunityRankingsResourcesPublish Guide
Plugin source
Back to catalog

sergey-akhalkov /

sergey-akhalkov/dsh-setup

Topic repository only

A portable DeepSeek Harness setup: sixteen plugins that make an agent work in small verified steps (second-model review, loop recovery, lazy MCP, local memory), plus a one-command installer.

★ 2 Stars0 Forks1 IssuesN/A Community rating0 Confirmed installs
View on GitHub
READMESource: main@f7cb2afe

dsh-setup

validate DeepSeek Harness 0.2.2-rc.1 License: MIT

A portable setup for DeepSeek Harness (DSH). Sixteen plugin bundles add rules, skills, MCP access, second-model review, recovery guards, and memory to an agent session; @dsh-setup/setup mounts all of them with working defaults, and one PowerShell command reproduces the setup on another machine.

What changes in a session:

  • The smallest complete change, with scope confirmed before it grows.
  • A fix is proven by its original failure instead of asserted.
  • Risky work comes back from a second model as a diff and a test report.
  • A runaway loop or a stranded goal ends with a result or a named blocker.
  • Context goes to the task: MCP tool schemas load on first use, bulky shell output is routed.
  • The repository is remembered between sessions, locally.

Every plugin is also installable on its own; use either the full setup or the individual plugins for a given plugin, never both, because both mount the same row.

Install or update

Install individual plugins from Discover

The setup and every plugin are listed in the published market index:

https://sergey-akhalkov.github.io/dsh-setup/index.json

Enable Custom plugin markets (this adds the Discover tab), open Discover, and add that URL as an index source. The index pins every listed version to a reviewed commit; install either the full setup or individual plugins for a given plugin, never both, because both mount the same row. See Manual install for the full steps.

Installing the dsh-setup entry from Discover mounts the setup and its plugin rows, but it does not enable the other shipped DSH bundles (Agent Teams, tool worktree, Ralph, session search, terminal). Enable them in Plugins, or use the full-setup command below. See Shipped DSH bundles.

Install the full setup with PowerShell

Use this on Windows when you want the complete setup in one command. Run it from an ordinary PowerShell window, not from a terminal hosted by DeepSeek Harness. Start DeepSeek Harness once and quit it first so the desktop profile exists.

irm https://raw.githubusercontent.com/sergey-akhalkov/dsh-setup/main/install.ps1 | iex

The installer prints its plan, closes DeepSeek Harness, backs up the profile manifest and patch, enables the shipped DSH bundles, installs or updates @dsh-setup/setup from the newest release, verifies the result, provisions the local Hindsight runtime when uv and a DeepSeek key are available, and starts the app again. Rerun the same command to update; when the newest release is already installed, it changes nothing, and a configured Hindsight runtime is skipped without prompting.

For -DryRun, a different profile, or a pinned version, download the script and run the file:

$installer = Join-Path $env:TEMP 'dsh-setup-install.ps1'
Invoke-WebRequest https://raw.githubusercontent.com/sergey-akhalkov/dsh-setup/main/install.ps1 -OutFile $installer
& $installer -DryRun

All options, the manual install, and the update behavior are in docs/install.md.

Dependencies

A missing optional runtime does not break startup: the affected adapter logs a failed connection, a skill fails only when called, delegation reports the missing target, and everything else keeps working. Install the runtimes for the plugins you want.

Dependency Needed for Install
DeepSeek Harness Desktop 0.2.2-rc.1 Running any plugin; the installer refuses an unsupported version Download DeepSeek Harness
PowerShell 5.1+ (7 recommended) The one-command installer Install PowerShell
Git The installer fetches the setup through a commit-pinned git spec Download Git
Node.js with npm Installing the two global CLI prerequisites below Download Node.js
uv (includes uvx) Serena MCP, every Codex delegation run, and the local Hindsight runtime Install uv
@nuphus/nuphus-mcp Nuphus MCP desktop and browser automation npm install -g @nuphus/nuphus-mcp
@fission-ai/openspec@1.15.0 The six OpenSpec skills npm install -g @fission-ai/openspec@1.15.0
rtk The RTK routing rule and skill Install RTK
Codex CLI (bundled) The Codex subagent and the default delegation target No separate install: the bundle ships @openai/codex 0.164.0; log in once with the command codex_subagent_status prints

What the setup installs

All sixteen plugin bundles are mounted and enabled by default, and the setup applies owner defaults: the deepseek-flash model at max reasoning effort, the dark theme, detailed transcript and usage display, and the cordis agent preset. A profile's own cordis.patch.yml still outranks these defaults.

"Always-on" contributions are a system-prompt section paid on every model step; on-demand skills are listed in the catalog and load only when invoked.

Plugin bundles

Plugin Enabled contribution What it gives you
@dsh-setup/plugin-ponytail Always-on ruleset + ponytail skill Minimal-code discipline: the smallest complete change, no unrequested abstractions, a reply a busy person reads once. Vendored upstream prompt (MIT, ~0.7k tokens per step); alwaysOn: false keeps the skill and drops the section.
@dsh-setup/plugin-working-principles Always-on working contract + working-principles, verify-change, instruction-budget, and interface-design skills Keeps scope with the user, demands evidence for claims, makes dissent concrete, and turns failures into recovery instead of retries; verify-change proves a fix by its original trigger; interface-design applies the first-time interface test across UI, CLI, API, and agent-facing definitions. ~1.5k tokens per step.
@dsh-setup/plugin-rtk Always-on routing rule + rtk skill Routes bulky git, test, build, and search output through RTK, cutting shell-output tokens while keeping errors and exit codes; native commands remain the fallback. ~0.2k tokens per step; needs the rtk CLI.
@dsh-setup/plugin-serena Always-on large-file/cross-file refactor trigger + serena-discipline skill Symbol-level navigation and editing with exact callers instead of whole-file passes, limited to large-file and cross-file refactor work; ordinary reads and text stay on grep/glob/read/edit. The setup configures the serena lazy server, so reach it with mcp_tools/mcp_call; no static adapter row is mounted. Needs uvx; ~0.2k tokens per step.
@dsh-setup/plugin-nuphus-mcp Nuphus MCP lazy server guidance + nuphus-discipline skill Desktop and browser control for tasks that need the real UI: windows, keyboard and mouse, local OCR and vision, Chrome over CDP. The setup configures the nuphus-mcp lazy server: reach it with mcp_tools/mcp_call; no static adapter row is mounted. Machine writes require explicit confirmation by default. Needs the nuphus-mcp binary.
@dsh-setup/plugin-mcp-lazy mcp_tools and mcp_call proxy tools over dormant stdio MCP servers Keeps configured stdio servers from paying their tool-schema cost in every request: a server starts on its first mcp_tools/mcp_call, is reused for the session, and is terminated on unload. The model-facing tool set stays at exactly two tools. The setup configures nuphus-mcp and serena as its lazy servers and mounts no static adapter row. Text results only, stdio only; a server must not also be mounted as a static adapter row (a call fails with a one-home error if it is). See lazy-mcp.
@dsh-setup/plugin-openspec Six on-demand OpenSpec skills Spec-driven changes in every session: propose, explore, apply, update, sync specs, archive. The skills ship here; specs live in each workspace's openspec/. Needs the openspec CLI.
@dsh-setup/plugin-rust rust skill (on demand) MSRV-correct standard-library use, the narrowest cargo check/test/clippy command, and diagnosis for slow builds, slow tests, and growing target directories. No per-step cost.
@dsh-setup/plugin-answer-page answer-me-with-html skill (on demand) Turns a complex answer into one self-contained HTML page rendered offline inside the workspace: the model writes a compact Markdown draft and the bundled answer-me-with-html 0.5.10 CLI (MIT, by QingYunA) owns layout, diagrams, and styling. Renders only on request or for a genuinely diagram-shaped explanation; the page stays under .tmp/answers/ in the workspace and is presented as a file. No per-step cost. See answer-pages.
@dsh-setup/plugin-delegation-rules delegate, task_progress, delegation_status tools, turn-stopping gate, skill Structured second-model reviews, research, and isolated worktree patches with digest-bound records; the strict gate stops a turn until a task is reviewed or a progress note explains the open work. Defaults target Codex; disabling the row removes the tools and gate.
@dsh-setup/plugin-loop-guard Host stream guard (no per-step prompt cost) Cuts repeated streamed reasoning early, steers a simple recovery notice into the next step, and escalates through bounded turn resets to a clear stop; enabled by default and subagents excluded. See loop-guard.
@dsh-setup/plugin-goal-continuity Host goal supervisor + block review + work_checkpoint pause gate (no always-on prompt section) Re-arms an active goal after recoverable harness stops: resumes a loop-guard reset, prevents the driver's pause, extends the round cap within a bounded budget, and blocks with a visible continuity-stopped reason on real stops. Every proposed blocked update is reviewed by a read-only goal-reviewer preset subagent before it takes effect: a denial continues with the next action, a verified blocker names the exact user-only action, and verdicts/grants are recorded locally; the reviewer preset stays hidden from the mode roster. See goal-continuity. Restart resume is off by default and resumeAfterUserStop defaults to non-goal; disabling the row restores stock behavior.
@dsh-setup/plugin-hindsight Native Hindsight memory + one-click local bootstrap Reuses the official native DSH plugin for automatic recall, retention, knowledge pages, and hindsight_* tools; defaults to a local hindsight-embed daemon at 127.0.0.1:8888 and autostarts it when unhealthy. The one-command installer provisions the runtime after the profile phase; setup-hindsight.ps1 runs only that phase. Disable the row or set autostart: false to opt out. See Hindsight.
@dsh-setup/plugin-codex-subagent Codex subagent provider with a pinned bundled CLI A second model for reviews, blockers, research, and bug-fix patches; runs @openai/codex 0.164.0, reuses your ~/.codex login, and pins its Serena MCP setup. One login is required; codex_subagent_status reports the command and the resolved CLI.
@dsh-setup/plugin-admin-indicator Desktop administrator label + shortcut Shows whether the Desktop app is elevated with the visible Standard User/Administrator badge next to the composer permission control; the menu creates an administrator desktop shortcut and shows a native success toast. The restart action is behind the default-off restart flag pending live UAC verification.
@dsh-setup/plugin-observability Observability dashboard: plugin health and signal counts derived from session logs (no second log) Answers whether a rule actually fired and a plugin is still working: a health verdict and KPI strip, an activity/error chart, plugin health cards grouped by attention with status words (active, error, not loaded, silent), tool distribution, and details on demand. The shipped per-plugin signal map covers every plugin; guidance rows name prompt-only plugins, and the blind spots are stated rather than hidden. Reads the session logs on demand and keeps nothing. See plugins.
@dsh-setup/plugin-verification-gate Host turn-close gate (no per-step prompt cost) Refuses to close a code-changing turn until one configured check command has succeeded after the last edit/write; inert with no commands, capped after maxBlocks (default 8) consecutive refusals, and only the configured command (optionally prefixed by one rtk token) satisfies it. See verification-gate.

@dsh-setup/setup is the meta-bundle: one install mounts all sixteen plugins and applies the owner defaults, and its publishable files are vendored so a git install needs no build step.

Shipped DSH bundles

The installer also enables the optional DSH bundles the setup expects; the setup bundle does not duplicate their rows:

  • Agent Teams (@deepseek-ai/dsh-experimental-agent-team-profile) — durable teammates, a shared task board, and isolated write scopes for parallel work.
  • Tool worktree (@deepseek-ai/dsh-experimental-tool-worktree) — per-agent git worktrees, so separate agents write without touching the same checkout.
  • Ralph (@deepseek-ai/dsh-experimental-ralph-bundle) — fresh-agent iteration rounds toward one immutable objective.
  • Custom plugin markets (@deepseek-ai/dsh-experimental-plugin-market-bundle) — the Discover tab and external market-index sources.
  • Session search (@deepseek-ai/dsh-experimental-session-search) — search across prior sessions.
  • Terminal (@deepseek-ai/dsh-experimental-terminal-bundle) — persistent terminal sessions for REPL and interactive work.

Development

Requires Node.js >=24.18.0 <25 and the pinned pnpm 11.28.5 (install pnpm).

pnpm install
pnpm check        # typecheck, lint, build, tests
pnpm validate     # full local gate, including generated-file and installer checks
pnpm dogfood      # Windows: verify install.ps1 against a copy of this machine's profile

Generated files are committed: the setup patch, lib/ output, and client.js bundles are built ahead of time so bundles install from git without an install-time build step. Run pnpm gen after changing a composition fragment and pnpm build after changing TypeScript sources; CI rejects a commit whose generated files do not match their sources.

Documentation

  • Install, update, manual install
  • Plugin development
  • Setup composition and the generated bundle
  • Releases, Pages, and the market index
  • Delegation rules and the Codex subagent
  • Plugin notes: Ponytail, working principles, RTK, Rust, Serena, OpenSpec, Loop guard
  • Repository operating guide: AGENTS.md

Security

Installing a bundle runs its code with the permissions of the Harness process. Market entries pin an exact git commit so the reviewed source and the running code are the same revision. Third-party prompts and skills are vendored with provenance locks (ponytail.lock.json, openspec.lock.json) that offline tests verify. The installer closes DeepSeek Harness while it writes the profile, saves a manifest backup, and changes only the bundle selection and installed packages — never the profile's own patch. This repository contains no credentials, sessions, or machine-specific paths.

License

MIT, copyright (c) 2026 Sergey Akhalkov. It covers the original code and documentation in this repository. Vendored third-party content keeps its upstream MIT license and attribution, with the upstream license shipped beside the files: the Ponytail prompt in packages/plugin-ponytail/assets/ponytail/, the OpenSpec skills in packages/plugin-openspec/assets/openspec/, and the Codex subagent provider in packages/plugin-codex-subagent/.

—/ 5

No ratings yet

Manifest verification required

Commit f7cb2afeb3b2

Community comments

No comments yet. Be the first to write one.

DSH HUB

A community index for DSH plugins. Not an official GitHub or DeepSeek AI product.

CommunityResourcesAPIAbout