dsh-pristine

August 16, 2026 · View on GitHub

A DeepSeek Harness agent preset that guarantees the first model request of every fresh session runs in a pure Minimal prompt state — zero tools, no baseline or skill injections — then processes the real prompt normally with the full Standard catalog.

This is a community project. It is not an official DeepSeek preset and is not affiliated with or endorsed by DeepSeek.

Features

Pure Minimal first request, by construction. The session's first step is replaced with a visible warmup round that carries only the fixed Minimal system prompt (You are a helpful software engineer assistant.) and NO tools. AGENTS.md/CLAUDE.md baselines, the skill catalog, and skill content never reach that request.
Zero-tool warmup. Even purer than a real Minimal session: the two tool schemas that Minimal normally ships are absent from the warmup request.
Injection shield. The replacement runs outermost on the pre-step waterfall, ahead of the instruction and skill injections, so those injections are discarded for the warmup request instead of polluting it.
Normal flow afterwards. The user's real first prompt is deferred by one turn and is then processed normally with the full Standard catalog, all instructions, and all skills.
Minimal shell and editor stack. The full catalog exposes Minimal's persistent bash (identical name, description, and schema) and str_replace_editor, while keeping the Standard file tools (read/write/edit/glob/grep).
Windows shell choice. On Windows, persistent Git Bash is the default (Git Bash requires Full access mode), with WSL and pwsh as explicit choices and a fail-soft fallback to pwsh.
Subagent coverage. Spawn-backed children warm up like the main session; fork children skip; subagents: false opts all delegated children out.
Fail-soft and idempotent. If the model route cannot be resolved, the first step is rejected, or the run is aborted, the warmup is skipped. Resumed or reloaded sessions never warm up again.
Offline and telemetry-free. The plugin performs no network requests and adds no telemetry.

Why

DeepSeek V4 Pro reaches its capability ceiling only under a Minimal prompt. On real projects that state is hard to hold: AGENTS.md/CLAUDE.md baselines, the skill catalog, and skill content are injected into the first request, so a nominally "minimal" first request is not actually minimal, and the capability-maximizing state is lost.

This preset holds the state by construction. The session's very first step is replaced with a warmup round whose request carries ONLY the fixed Minimal system prompt and NO tools at all — even purer than a real Minimal session, whose two tool schemas are absent from this request. The replacement runs outermost on the pre-step waterfall, ahead of the instruction and skill injections, so those injections are discarded for this request instead of polluting it. The user's first prompt is deferred to the next turn and is then processed normally: full Standard catalog, all instructions and skills.

How it works

On the first input of a fresh session, the first step becomes a warmup round instead of answering the user.
The replacement runs outermost on the pre-step waterfall, ahead of the instruction and skill injections, so those injections are discarded for this request instead of polluting it.
The warmup request contains only the Minimal fixed system prompt (You are a helpful software engineer assistant.) and NO tools. No AGENTS.md/CLAUDE.md baseline, no skill catalog, no skill content.
The warmup message is a bare round-framing sentence: "This round is a test. Tools are not open yet; all tools will open next round." It names no goal, tool, or command, so the message itself disturbs the pure Minimal request as little as possible.
The warmup turn is visible in the trajectory like any other turn.
The real first prompt is then processed normally with the full Standard catalog. The shell in the full catalog is Minimal's persistent bash (identical name, description, and schema). On Windows the default backend is Git Bash, with WSL and pwsh available as explicit choices, and str_replace_editor joins the Standard file tools (read/write/edit/glob/grep are kept).

Warmup coverage

Session	Warmup?	Why
Main fresh session	Yes	Its first request would otherwise carry the baseline and skill injections.
Spawn child (`subagent` tool, workflow `agent()` calls, ralph rounds)	Yes	Each starts from an empty log and faces the same injections.
Fork child (`subagent_fork`)	No	It inherits the parent's completed turns, so its first request can never be pure Minimal. A fork child with an empty seed is indistinguishable from spawn and warms up — harmless.
Resumed or reloaded session	No	It has already recorded a `request/header`.
External CLI subagent (codex, claude-code)	Not covered	It does not run the agent loop.

Configuration

Delegated sessions

The warmup is enabled by default for spawn-backed children. To skip it for every delegated child — useful for bulk fan-out that should not pay one extra turn per child — edit the installed preset's agent.cordis.yml:

- id: warmup-bootstrap
  name: ./warmup-bootstrap.mjs
  config:
    subagents: false

Windows shell

On Windows, Pristine defaults to a persistent bash backed by Git Bash. Git Bash only takes effect while the session is in Full access mode — switch the session to Full access before relying on the default shell. To choose a different shell, edit the installed preset's agent.cordis.yml under the windows-shell-bootstrap config:

- id: windows-shell-bootstrap
  name: ./windows-shell.mjs
  config:
    windowsShell: git-bash   # git-bash (default), wsl, or pwsh
    gitBashPath: null        # optional absolute path to bash.exe
    wslDistro: null          # optional WSL distro name, e.g. Ubuntu

git-bash auto-detects common Git for Windows locations; gitBashPath overrides detection when Git is installed somewhere non-standard.
wsl uses wsl.exe; wslDistro selects a non-default distribution and must name a distro already installed in WSL.
While a bash variant is active, the model-visible catalog exposes only bash and hides pwsh.
If the selected bash variant is unavailable or the configured value is invalid, Pristine logs a warning and falls back to pwsh so the session remains usable.

Compatibility

Developed and tested against:

DeepSeek Harness 0.1.0-rc.6
Node.js 22 on Linux

DeepSeek Harness is currently a developer preview and explicitly permits breaking changes. This preset is a snapshot of the Standard composition with Minimal's shell and editor stack swapped in, so review upstream changes before using it with a newer release.

Install

Scripted install

Linux/macOS:

./install.sh           # snapshot copy (default)
./install.sh --link    # symlink preset/ so git pull updates apply immediately

Windows (PowerShell):

.\install.ps1          # snapshot copy (default)
.\install.ps1 -Link    # symlink (requires Developer Mode)

Both scripts resolve the preset root from DSH_HOME (falling back to ~/.dsh or %USERPROFILE%\.dsh), install under the id pristine, leave an existing installation untouched (an identical install is reported as already installed), and verify the installed files. Run with --help or -Help to list all options. If PowerShell's execution policy blocks the script, use powershell -ExecutionPolicy Bypass -File .\install.ps1.

Uninstall

Remove the preset with ./uninstall.sh (Linux/macOS) or .\uninstall.ps1 (Windows). Both scripts prompt for confirmation; use --yes or -Yes to skip the prompt.

Manual install

Clone this repository, then copy the entire preset directory into the user preset root under the id pristine.

PowerShell:

$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\pristine'
if (Test-Path -LiteralPath $target) { throw "Preset already exists: $target" }
New-Item -ItemType Directory -Force -Path (Split-Path -Parent $target) | Out-Null
Copy-Item -Recurse -LiteralPath '.\preset' -Destination $target

Linux/macOS:

dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets"
test ! -e "$dsh_home/.agent-presets/pristine"
cp -R preset "$dsh_home/.agent-presets/pristine"

After installing

Fully restart DeepSeek Harness, create a blank session, and select Pristine. Do not switch an active session from a different preset.

Verify

The first visible assistant turn is the warmup: its reply is a short acknowledgment, not an answer to the user's prompt.
Export the session JSONL and inspect request/header events: the first header carries NO tools; the next header carries the full Standard catalog.
Spawn a subagent (e.g. via the subagent tool) and export its session JSONL: its first header also carries NO tools.

Run the local zero-dependency tests with:

npm test

Behavior notes

Failure handling. The warmup is fail-soft: if the model route cannot be resolved, the first step is rejected, or the run is aborted, the warmup is skipped and the real input proceeds unchanged.
Cost and visibility. The warmup consumes one real turn (and its tokens) and is visible in the trajectory. Every fresh session pays it — the main session and each spawn-backed subagent (subagents: false opts subagents out).
One warmup per fresh session. Sessions that already recorded a request/header (resumed or reloaded sessions) do not warm up again.
Catalog scope. The catalog is narrowed only while a warmup is pending: while pending it is narrowed to zero tools, and the following turn gets the full catalog back.
Windows fallback. On Windows, the selected Git Bash/WSL shell is fail-soft: if it is missing or misconfigured, the preset logs a warning and falls back to pwsh.
Trust. The preset has the same trust level as shell access. Review its files before installation.
Privacy. The plugin performs no network requests and adds no telemetry.

Official ecosystem guidance

DeepSeek currently asks community plugin authors to publish plugins in their own GitHub projects and add the dsh-plugin repository topic for discovery. The official repository does not currently accept external pull requests and does not mandate a community repository template. See the official CONTRIBUTING.md.

Acknowledgments

The core insight behind this preset — that DeepSeek V4 Pro reaches its capability ceiling only under the Minimal prompt state — comes from dsh-anchored-standard. This preset builds on that work and closes the gap it left on real projects: the first model request is now guaranteed pure Minimal, because AGENTS.md/CLAUDE.md baselines and skill injections are discarded from the warmup round instead of polluting it.

License

MIT. preset/agent.cordis.yml is derived from the DeepSeek Harness Standard preset; the original DeepSeek copyright and MIT notice are retained in NOTICE.