@chaggle/dsh-powershell-check
A native DeepSeek Harness plugin that gates every pwsh tool call against the PowerShell pitfalls documented in the blog post PowerShell 实战踩坑大全 (GBK console mojibake, $var: parsing, PS 5.1 ternary, double-quoted variable expansion, JS escaping, Start-Process quoting and sandbox traps, -FeatureName arrays, DISM verbs, RestoreHealth source versions, CDN downloads, npm.cmd suffix, &&/|| chains), and bundles the powershell-check skill.
Installable as a profile plugin: dsh plugin --profile <name> add @chaggle/dsh-powershell-check (or mount the row directly — see Install).
Features
- Automatic gate — subscribes to the official
tools/pre-execute interception point; every pwsh call is statically checked before execution. Blocking violations (R2–R11) return a deny whose reason IS the fix guidance; the advisory R1 passes through. warn mode only logs.
- Bundled skill — exposes
powershell-check through ctx.skills.registerProvider (the dsh-skill-badge pattern), visible and loadable in every session catalog.
- Bilingual — rule text, CLI output (
--lang en|zh), deny reasons, and docs ship in English and Simplified Chinese.
- Single rule source — the rules engine (
src/checker.ts) is shared by the gate and the CLI; update once, both follow.
- Self-test —
--selftest runs a 60-case positive/negative battery (R1–R19, AI-generated-script focused).
Install
As a profile plugin (recommended)
The package declares dsh.bundle.patch and ships its own cordis.patch.yml, so it is installable per profile:
dsh plugin --profile <name> add @chaggle/dsh-powershell-check
or add the row to your profile layer $DSH_HOME/profiles/<name>/cordis.patch.yml:
- insert:
- id: dsh-powershell-check
name: @chaggle/dsh-powershell-check
config:
mode: deny # deny | warn
lang: zh # zh | en
If the harness was launched from a checkout, either publish the package or point the row at a local clone:
git clone https://github.com/chaggle/dsh-powershell-check.git
# then junction/symlink it into $DSH_HOME/profiles/node_modules/@chaggle/dsh-powershell-check
User patch layers are watched: the change hot-applies to a running dsh web instance (transactional HMR) without a restart.
Skill only (no gate)
The repository root is a skill bundle (SKILL.md + scripts/); clone it into any skill root:
git clone https://github.com/chaggle/dsh-powershell-check.git "$HOME/.dsh/skills/powershell-check"
Configuration
| Key | Default | Meaning |
|---|
mode | deny | deny blocks pwsh calls with blocking violations; warn logs and allows |
lang | zh | Language of deny reasons and warn logs: zh or en |
analyzer | builtin | builtin: bundled R1–R19 rules only. psscriptanalyzer: additionally deep-checks every pwsh command with the official PSScriptAnalyzer (install the module on the host; degrades to builtin when missing). Error/ParseError findings deny, warnings are logged |
CLI
node scripts/check-pwsh.mjs -- "command text" [--lang en]
Get-Content fix.ps1 -Raw | node scripts/check-pwsh.mjs - [--lang en]
node scripts/check-pwsh.mjs --selftest
Exit codes: 0 = PASS, 1 = FAIL (violations listed with fixes), 2 = usage error.
Rules
| Rule | Level | Detects | Blog section |
|---|
| R1 | advisory | wsl/Windows feature queries without chcp 65001 | §1-1-3 |
| R2 | blocking | $var: parsed as drive-qualified syntax | §1-1 |
| R3 | blocking | ) ? ... ternary shape (PS 5.1) | §1-2 |
| R4 | blocking | $WORD followed by ./`` inside double quotes (path trap) | §1-1-1 |
| R5 | blocking | Start-Process wrapping an external command | §1-2-1 |
| R6 | blocking | -FeatureName A, B array form | §1-1-1-1 |
| R7 | blocking | /Dismount-Image verb | §1-1-1-2 |
| R8 | blocking | RestoreHealth with a newer source | §1-1-1-3 |
| R9 | blocking | curl -L combined with -C - | §1-2-4 |
| R10 | blocking | bare npm/npx/pnpm without the .cmd suffix | §1-2-2 |
| R11 | blocking | && / ` | |
| R12 | advisory | ConvertTo-Json without -Depth (default 2 truncates nested data) | §5-1 |
| R13 | advisory | $_ inside a foreach ($x in ...) loop | §5-2 |
| R14 | blocking | single = as a comparison inside if/while | §1-4 |
| R15 | advisory | PS 7+ only syntax (-AsHashtable/-Parallel/-AsByteStream/??/?.) | §1-5 |
| R16 | advisory | cmd-style commands / env syntax |
PSScriptAnalyzer deep check
Set analyzer: psscriptanalyzer to add the official PowerShell static analyzer on top of the builtin rules. The plugin runs Invoke-ScriptAnalyzer against the command text (as a temp .ps1) through the same ctx.shell seam the harness hooks bridges use; the probe raises the process-scope execution policy to Bypass first (the harness starts pwsh under Restricted). Findings with severity Error/ParseError deny the call with the analyzer messages; warnings are logged and allowed.
# one-time, on the host (PowerShell 5.1 or 7):
Install-Module PSScriptAnalyzer -Scope CurrentUser -Force
Trade-offs: each deep check spawns an analyzer pass (module load ~1–3 s), so enable it only when the extra coverage is worth the latency; when the module is absent the plugin logs once and falls back to the builtin rules. Empirical notes from this project: PSSA does not flag &&/|| on a 5.1 host (the builtin R11 covers that) and serializes Severity as a numeric enum (the parser handles both forms).
How it works (official extension points)
The harness extension surface is its typed interception points: a "native hook" is an ordinary Cordis plugin subscribing to canonical lifecycle events and returning typed decisions — no external hooks bridge, no hook/* log, no subprocess boundary. This plugin uses two entries:
ctx.on(tools/pre-execute, (exec, next) => PreToolDecision) — the pre-execution waterfall gate;ctx.skills.registerProvider(...) — contributes the bundled skill to the registry.
Model Experience
Request context and condition
What the model sees
Nothing is injected into prompts by this plugin. The skill is exposed through the standard session skill catalog (powershell-check, description above) and can be loaded with the skill tool. When the gate denies a pwsh call, the model sees the deny reason in the tool error result — that reason is the fix guidance generated by formatHits.
Token effect
Zero direct token effect outside tool-error feedback: no prompt text is added or rewritten. The deny reason replaces a would-be tool result with a bounded error payload.
KV Cache effect
The plugin publishes no system-reminder or catalog text of its own; the skill description rides the session catalog produced by the skill consumer, so prompt-prefix reuse is unaffected. Deny reasons are per-call error results and do not change the request prefix.
Known Limitations and Deferred Work
- Static heuristics — rules are pattern-based; R1 is advisory by design and R4 may flag intended variable expansion (the fix text says when to ignore).
- Live reload — code changes require the running harness to re-import the plugin (user-patch rows hot-reload; the module cache reloads on row replacement).
- Rule count — R10/R11 cover the ExecutionPolicy and
&&/|| traps; new pitfalls should be added to src/checker.ts with selftest cases, then mirrored in the blog post.
License
MIT — see LICENSE.
