@agent-shells/powershell-skills

powershell-skills

PowerShell skills and guardrails for reliable Windows command execution by AI agents.

General-purpose coding agents often run Windows and PowerShell less smoothly than Linux or macOS shells. This project gives agents a small, tested runtime aid: agent-facing skills, safe command shapes, compatibility checks, and helper scripts that reduce common Windows command failures.

Installation

Install from npm:

npm install -g @agent-shells/powershell-skills
powershell-skills install all
powershell-skills doctor

Update explicitly:

powershell-skills update

Install only one adapter:

powershell-skills install codex
powershell-skills install claude-code

GitHub install is still supported for unreleased branches or tags:

npm install -g github:agent-shells/powershell-skills#v0.5.1

Features

• Codex adapter for user-level skill discovery under ~/.codex/skills.
• Claude Code adapter for personal skill discovery under ~/.claude/skills.
• powershell-skills CLI for install, update, and doctor workflows.
• Shared execution contract, pattern catalog, and sanitized failure corpus.
• PowerShell helper scripts for path resolution, command discovery, failure classification, external application execution, and read-only cmdlet execution.
• Invoke-AgentPowerShell.ps1 for structured, allowlisted, read-only cmdlets.
• Windows PowerShell 5.1 and PowerShell 7 verification through GitHub Actions CI.

Triggering

The adapters are written for agents, not for humans to prefix every command manually.

• Codex discovers powershell-command-runner from ~/.codex/skills.
• Claude Code discovers powershell-command-runner from ~/.claude/skills.
• Both adapters describe Windows, PowerShell, path handling, encoding, filesystem, timeout, and command-failure tasks as trigger cases.

If a current agent session has not refreshed its skill index, restart the agent or start a new session. Details are in adapter behavior.

Compatibility

The project is tested on:

• Windows PowerShell 5.1 through powershell.exe.
• PowerShell 7 through pwsh.
• Node.js 18+ for the npm CLI wrapper.

The npm layer is distribution and install UX. PowerShell scripts remain the compatibility source of truth. See compatibility and verification.

Commands

powershell-skills install codex
powershell-skills install claude-code
powershell-skills install all
powershell-skills doctor
powershell-skills update

Use powershell-skills doctor after install or update to check package files, PowerShell hosts, npm, and installed skill locations.

Verification

Run local checks:

npm test
npm pack --dry-run
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\verify-v0.1.ps1
pwsh -NoProfile -ExecutionPolicy Bypass -File .\scripts\verify-v0.1.ps1 -PowerShellExe pwsh

Expected result:

[OK] npm CLI smoke tests passed
[OK] release verification passed

Docs

Topic	Document
Agent adapters and auto-trigger behavior	docs/adapters.md
Helper scripts and runners	docs/helpers.md
PowerShell 5.1 / 7 compatibility	docs/compatibility.md
Architecture and repository layout	docs/architecture.md
Security model and contribution boundaries	docs/security-model.md
Field-test notes	docs/field-tests/README.md

Release Notes

• v0.1: Codex skill foundation.
• v0.2: PowerShell 5.1 / 7 compatibility matrix and CI.
• v0.3: Claude Code adapter.
• v0.4: npm distribution, update, doctor, and install UX.
• v0.5: read-only PowerShell cmdlet helper.
• v0.5.1: README and documentation polish.

Contributing and Security

• Contributing guide
• Security policy
• Changelog

This project accepts normal reviewed issues and pull requests. It does not include automatic telemetry, automatic failure upload, periodic collection, or upload of user command context.

Current Limits

• Supported adapters are Codex and Claude Code.
• Updates are explicit; there is no background auto-update service.
• Invoke-AgentCommand.ps1 runs external applications only.
• Invoke-AgentPowerShell.ps1 is read-only and allowlist-based; it is not a general PowerShell script runner.
• Destructive command execution is not automated and requires validation outside the runner.
• The failure corpus accepts only maintainer-reviewed, sanitized, minimized cases.