ArmyWas / dsh-plugin-reducer

dsh-plugin-reducer

简体中文

Your DeepSeek Harness profile is broken. Which plugins are actually required to reproduce it?

dsh-plugin-reducer is an unofficial, external diagnostic CLI that finds a 1-minimal failure-inducing set of out-of-tree bundles in a DeepSeek Harness profile. It tests disposable shadow profiles, so it does not rewrite the real profile while searching.

It catches interaction failures that "disable plugins one by one" misses. If A works, B works, and A+B fails, the result is {A, B}.

Early preview: tested against @deepseek-ai/dsh@0.1.0-rc.7 on Windows with Node.js 24. The repository CI also targets Windows, macOS, and Linux on the Node.js versions supported by Harness.

Found a real profile failure? Open a redacted field report.

Why this exists

A profile can contain dozens of bundles. Startup guards can recover a broken profile, doctors can identify common problems, and repro tools can collect a session. None of those answers the narrower question:

What is the smallest plugin set I must give an author so this exact failure still happens?

The reducer turns a large, private profile into a small, evidence-backed plugin set and a scrubbed JSON report suitable for an issue.

Quick start

From a source checkout:

git clone https://github.com/ArmyWas/dsh-plugin-reducer.git
cd dsh-plugin-reducer
npm install
npm test
npm link
dsh-plugin-reducer --profile web --probe web --report reducer-report.json

After an npm release, the same run can use npx:

npx dsh-plugin-reducer@latest --profile web --probe web \
  --report reducer-report.json

If Harness lives outside PATH, point to it explicitly:

dsh-plugin-reducer --dsh /path/to/dsh --profile web --probe web

The tool reads DSH_HOME from the environment, or accepts --dsh-home.

List candidates

Before spending probe runs, see exactly which out-of-tree bundles the reducer would consider. --list-candidates reads the profile manifest, prints one bundle name per line, and exits 0:

dsh-plugin-reducer --list-candidates --profile web

plugin-a
plugin-b
plugin-c

This is a read-only listing: it does not create a shadow DSH_HOME, does not run any probe, and does not require the dsh executable to be installed or on PATH. Probe-only options and a command after -- are rejected instead of being silently ignored.

What a run does

1. Reads $DSH_HOME/profiles/<name>/package.json.
2. Treats names present in both dsh.profile.bundles and dependencies as reducible, out-of-tree candidates. Installation-owned bundles stay fixed.
3. Creates a fresh temporary shadow DSH_HOME for every probe attempt and links its profile to the existing node_modules tree. It installs nothing.
4. Runs the selected probe against subsets and complements using delta debugging.
5. Verifies the result by removing each remaining bundle once.
6. Optionally writes a secret- and path-scrubbed JSON report.

Example result:

Minimal failing set (2):
  - plugin-a
  - plugin-b
Verified 1-minimal: yes; 6 configuration(s) tested.

"1-minimal" means removing any one reported bundle stops reproducing the failure. It does not claim the set is the globally smallest possible set.

Probes

Probe	Pass condition	Best for
config (default)	dsh --dump-config exits 0	Manifest and layer-composition failures
web	A random loopback URL answers HTTP and survives the settle window	Plugin loading and startup crashes
custom command	The supplied command exits 0	A precise regression, test, or workflow

For a custom probe, put the command after --. The child receives the shadow home in DSH_HOME:

dsh-plugin-reducer --profile web --timeout 60000 -- node reproduce.mjs

The oracle should exit non-zero only when the target failure is reproduced; unrelated command errors are also classified as failures, so keep it specific.

For flaky failures, repeat every configuration. Mixed outcomes are marked unresolved and are never used as evidence to remove a plugin:

dsh-plugin-reducer --profile web --probe web --repeat 3 --max-trials 512

Run dsh-plugin-reducer --help for all options.

Add --keep-lab to retain a clean shadow home configured with the final minimal set for manual inspection. Probe-attempt labs are still removed.

Safety boundary

• The CLI writes candidate changes only inside a temporary shadow home.
• Every probe attempt receives a fresh shadow home, preventing state written by one tested configuration from contaminating the next.
• It fingerprints the source profile manifest and patch files before and after the run and records whether they changed externally.
• It reuses installed packages and never invokes a package manager or install script.
• It does not copy .env, sessions, storage, or workspace data into the lab.
• Reports replace known local roots and fallback absolute paths, and redact common token forms and secret-shaped fields.

This is a diagnostic isolation mechanism, not a security sandbox. Installed plugins and custom probes execute with the current user's permissions and may access the network, inherited environment variables, or files they could access in a normal Harness run. Review SECURITY.md before running untrusted code.

Scope and limits

• The failure must reproduce consistently in the shadow profile.
• Built-in bundles and profile/home patches remain fixed; if the empty out-of-tree baseline fails, the reducer stops with BASELINE_FAILS.
• Packages installed only as libraries, not listed as bundle layers, are not candidates.
• The Web probe tests startup readiness, not a complete interactive agent flow.
• The algorithm optimizes diagnostic effort and 1-minimality, not proof of a globally minimum set.

How this differs from existing projects

Project	Primary job	Relationship to this project
dsh-startup-guard	Preflight, recovery, snapshots, quarantine	Recover first; reduce a stable reproduction afterward
dsh-boot-guard	Loader-independent manual rescue and plugin skipping	Restore control; it does not search interaction sets
dsh-repro	Export a scrubbed session/command/git-diff reproduction	Capture session context; reducer minimizes profile bundles
dsh-builtin-toggles	Inspect and toggle built-in capabilities	Runtime inspection rather than failure-set reduction

The ecosystem review records the search and the decision boundary. This tool is designed to complement those projects, not replace them.

Project evidence and design

• Real Harness interaction test
• Real Harness dogfood report for --list-candidates
• Product brief
• Report JSON Schema
• Upstream RFC
• Community launch kit
• GitHub publication runbook
• Official Discussion draft
• v0.1.0 release notes
• v0.2.0 release notes
• Contributing

DeepSeek Harness is a trademark of its respective owner. This project is not affiliated with or endorsed by DeepSeek.