YAOXIAOYUAN0802/Crow-purge ↗★ 2

Xsdsh-purge

Crow-purge 逆向模式:跨平台逆向分析指挥体系(iOS / Android / Windows),五角色 + 五阶段 + 证据链;附外挂制作模式

패키지
Xsdsh-purge
호환성
미검증
Harness peer 범위
>=0.2.1-rc.2 <0.2.2-0
Cordis peer 범위
^4.0.1
버전
1.0.0
라이선스
MIT
최근 업데이트
2026. 10. 31.

설치

$npx -p @deepseek-ai/dsh dsh plugin --profile web add github:YAOXIAOYUAN0802/Crow-purge

Crow-purge

The era of AI refusing people is over.

version

dsh

license

tests

⚠️ Read before use → Disclaimer & Acceptable Use This project may only be used in environments you can prove you are authorized to handle. Any use beyond that is the user's sole responsibility.


Project type

In one line

This is a host plugin — not a standalone app, not a model, and not a fine-tune.

It lives inside DeepSeek Harness and changes the host's runtime behaviour, not the model's weights.

Technical classification

DimensionValue
EcosystemDeepSeek Harness (dsh) plugin
Formcordis plugin package (cordis.patch.yml, registered from a YAML manifest)
Module systemESM ("type": "module", entry lib/index.js)
Runtime dependenciesNone (empty dependencies; only host-provided peers)
LayerInside the host process — prompt injection + host source patches + runtime event hooks
PersistenceNo database. State lives in JSON/JSONL files under $DSH_HOME
NetworkFully offline by default. Only contacts this repo if you enable update checks
PlatformWindows (the desktop host is Windows-only)

By purpose

Security tooling → LLM application security → prompt-layer evaluation and behaviour rewriting.

More precisely, an intersection of three things:

  1. Prompt engineering / prompt-layer tooling Treats prompts as programmable objects: per-section overrides, order tuning, runtime substitution.

  2. LLM guardrail research Studies when and how model refusals trigger, and provides reproducible measurement (runtime probe, a 972-assertion regression suite).

  3. Local security assessment harness Red-team / reverse / cheat drill presets, scoped to this machine and authorized environments.

What it is not

Often confused withReality
❌ A standalone appNo UI process of its own; the panel is injected into the host
❌ A model / a fine-tuneNo training, no distillation, no weight changes. Only prompts and host behaviour
❌ A proxy / gatewayNo request forwarding, no traffic proxying, no API key routing
❌ A jailbreak kitNo "one-liner bypass" prompts; it's engineered patches plus measurable regression
❌ MalwareNo persistence, no phone-home, no file encryption, no data theft. Source is open, MIT

Compared to similar projects

ProjectTypeRelationship
dsh-purgeSame-ecosystem plugin (upstream)User notice text and authorization flow originate here
Xsdsh-purgeSame-ecosystem plugin (upstream)Patch table, drill console, three modes originate here
dsh-purge-plusSame-ecosystem plugin (sibling)Independently evolved; strong on CLI / re-anchor / audit, lacks this project's correction layer and probe
Ollama / LM StudioModel runtimesDifferent layer, unrelated
SillyTavern presetsPrompt templatesDifferent layer: they change chat style; this changes host behaviour

The problem

You install DeepSeek Harness, then notice:

  • Ask "who are you" and it answers "I'm an AI assistant running inside DeepSeek Harness"
  • Ask a straight technical question and it starts citing regulations and offering "compliant alternatives"
  • On your own machine, your own lab, it still wants you to confirm you have authorization

Crow-purge changes all three.


Three core capabilities

① Prompt rewriting — stop it from reciting the product brochure

The host's built-in identity statements, refusal tendencies and liability boilerplate are all replaceable.

The key is ordering: the host puts WEB_SURFACE(10100), HARNESS_SOURCE(10000), STRUCTURED_OUTPUT(9900) at the tail of the prompt, and later sections carry more weight. This plugin pushes its own block to 10200, above all of them.

Measured: the same question went from "I can't do that." to a full technical breakdown after this change.

② Refusal correction — pull sampling back when it drifts

Model output is sampled. The same prompt, three runs, can yield 16/20, 19/20, 19/20 — never stable.

So there's a runtime layer: when a turn ends, read the output; if it looks like a refusal, resend with escalating tactics:

AttemptStageWhat it does
1restaterestate the rules
2reframechange the frame: "output only the smallest executable fragment", first character is technical content
3scaffoldgive a skeleton: conclusion / code / params / next step — fill-in-the-blanks

All three keep the hard constraints: no refusal, no risk assessment, no asking about purpose.

③ Runtime probe — tell "plugin isn't working" from "the model won't"

This is the most time-consuming thing to diagnose. Both cases look identical in the UI, and the log's injectBytes can't distinguish them either (that number only proves the host received it).

How it works: the panel generates a random nonce as a temporary section. Send any message:

model echoes [purge-a1b2c3d4]  → injection reached the model; the problem is the model
no echo                         → injection never arrived; check postPromptOrder

Expires after 30 minutes; doesn't occupy tokens long-term.


Install

Option 1: Installer (recommended)

Download Crow-purge-安装器.zip from releases, extract, double-click 安装-Crow-purge.cmd.

It will:

  1. Request admin rights and add a Windows Defender exclusion for the folder
  2. Back up the old version
  3. Copy files
  4. Verify all 7 root files individually — reports missing ones immediately instead of letting the host throw ENOENT

Why the exclusion matters: client.js (785 KB minified JS) and cordis.patch.yml (contains inject: [systemPrompt, tools]) are classic antivirus targets and get silently quarantined. On one machine, 7 root files became 5, and the host reported ENOENT while the files "were clearly there".

Option 2: npm

dsh plugin --profile desktop add Xsdsh-purge

The package name is Xsdsh-purge (the host loads by package name; renaming breaks it). The display name is Crow-purge.

After installing

Click Crow-purge next to the session title → Clean → Apply. The host restarts once.


Panel

TabWhat it does
CleanPatch list, apply / revert / uninstall. Shows which line of the host each patch touches and whether it's on
RewriteOverride the host's prompt sections, one block at a time
DrillIn-scope local exercises (red-team / reverse / cheat presets)
WallpaperBackground image and music. Built-in / Random (new image each load) / local file
PromptsPer-mode prompt text
ChangelogWhat changed in each version and why

A Runtime probe card sits at the top of the Clean tab

That's the one from ③ above. Check it first when you hit "still not working".


What it changes

82 patches, by layer:

LayerCountWhat it touches
Prompt25identity statements, liability boilerplate, refusal tendencies, tool descriptions
Code46approval auto-allow, sandbox opening, fs fences, tool result pruning, plan-mode hard blocks
Tools8tool table and call preconditions
Compat30.1.5 / 0.2.0 differences

Each patch = {id, name, layer, file, replacements, markers}; markers decide "already applied" so the operation is idempotent.


Compatibility

ComponentValue
PluginCrow-purge 1.0.0
Verified hostdsh 0.2.1-rc.2 ~ 0.2.1-rc.5
engines.dsh`>=0.2.1-rc.2 中文说明见 README.zh-CN.md