Skill stocktake

Evidence-backed, resumable skill ecosystem auditing for the Codex app and CLI.

Install

npx -y skills add ToxicityRadius/skill-stocktake

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use when auditing, inventorying, comparing, consolidating, updating, retiring, or checking changes across user, repository, admin, system, or managed Codex skills, especially when activation, overlap, freshness, usage, dependencies, or resumable cached reviews matter.

SKILL.md

4.4 KB, as published. Nobody here has run it

Skill Stocktake

Audit skills without changing them. Require explicit approval before rewriting, merging, archiving, or removing an audited skill. Treat admin, system, and managed/plugin skills as read-only.

Prepare

Resolve the installation and project context. ProjectRoot is the repository root; CurrentWorkingDirectory determines nested skill and instruction scope. Prefer the cross-platform Python v4 engine.

$CodexHome = if ($env:CODEX_HOME) { $env:CODEX_HOME } else { Join-Path $HOME '.codex' }
$SkillRoot = Join-Path $CodexHome 'skills\skill-stocktake'
$GitRoot = & git rev-parse --show-toplevel 2>$null
$ProjectRoot = if ($LASTEXITCODE -eq 0 -and -not [string]::IsNullOrWhiteSpace($GitRoot)) { $GitRoot } else { (Get-Location).Path }
$CurrentWorkingDirectory = (Get-Location).Path
$StatePath = powershell.exe -NoProfile -ExecutionPolicy Bypass -File (Join-Path $SkillRoot 'scripts\Get-DefaultStatePath.ps1') -ProjectRoot $ProjectRoot

Confirm Python 3.10+ is available, then run python -m skill_stocktake doctor from the skill directory. The PowerShell entrypoints remain compatibility paths for v3-era automation.

Always read references/evaluation-rubric.md before assigning verdicts or constructing records.

Build the worklist

$WorkPath = Join-Path (Get-Location) 'skill-stocktake-work.json'
$RunPath = Join-Path (Get-Location) 'skill-stocktake-run.json'
$WorkJson = powershell.exe -NoProfile -ExecutionPolicy Bypass -File (Join-Path $SkillRoot 'scripts\Get-SkillDiff.ps1') `
  -StatePath $StatePath -ProjectRoot $ProjectRoot -CurrentWorkingDirectory $CurrentWorkingDirectory -UsageMode None
[System.IO.File]::WriteAllText($WorkPath, ($WorkJson -join [Environment]::NewLine))
powershell.exe -NoProfile -ExecutionPolicy Bypass -File (Join-Path $SkillRoot 'scripts\New-AuditRun.ps1') -WorklistPath $WorkPath -OutputPath $RunPath | Out-Null

Session-history analysis requires explicit user approval. Use Python --include-usage or PowerShell -UsageMode Sessions only after approval. Pass additional or symlink roots only when confirmed by the current runtime and explicitly trusted.

The inventory distinguishes confirmed, inferred, disabled, and excluded activation. Never call coverage complete when diagnostics report unreadable roots or inferred managed activation.

Follow suggested_mode:

full: evaluate every logical skill.
quick: evaluate only diff.pending_ids and carry valid unchanged records forward.
resume: continue the exact active run. If inventory or context changed, start a new full run and use Save-Results.ps1 -ReplaceActiveRun for its first save.

Evaluate

Inspect each selected skill's complete SKILL.md and only the resources needed to validate its claims. Apply every rubric dimension and map evidence to each concern, fail, or unknown dimension.

Verify changing platform claims with primary sources. Check applicable instruction precedence, activation state, metadata, tool dependencies, and reverse references. Treat usage as supporting evidence only; zero or unavailable usage never independently justifies removal.

Evaluate in batches of about 20. Update the run JSON with complete pending/evaluated lists after each batch. Use multiple agents only when the user explicitly requests them; the parent remains the sole state writer.

Save safely

powershell.exe -NoProfile -ExecutionPolicy Bypass -File (Join-Path $SkillRoot 'scripts\Save-Results.ps1') -StatePath $StatePath -EvaluationPath $RunPath

Schema version 3 validates identity, progress, verdicts, confidence, mapped evidence, proposal fields, freshness, and merge/retirement safeguards. Saving is locked and atomic. Never bypass validation.

Report

powershell.exe -NoProfile -ExecutionPolicy Bypass -File (Join-Path $SkillRoot 'scripts\Format-Report.ps1') -StatePath $StatePath -WorklistPath $WorkPath

Report coverage and diagnostics first, then summary counts, actionable or low-confidence results, the complete table, proposals, and uncertainty. Separate confirmed runtime facts from cache inference. Do not apply proposals during the audit.