GESH Lab · LLM tooling

Lint the skill before it taxes every turn.

Zero-dependency linter for Claude Code and Agent Skills. Frontmatter, progressive disclosure, dead refs, secrets, stale model IDs — plus a 22-skill NetOps pack for the change window.

Release v0.5.0 Install the CLI

89lint checks

22NetOps skills

0runtime deps

3profiles

2026Claude fields

What 0.5 adds

CLI and playground now agree. Pin the GitHub Action at @v0.5.0. Docs live here; the engine lives in one Python 3.8+ file.

2026 models

Opus 5, Sonnet 5, Fable 5, Mythos 5 are current. Opus 4.8 / Sonnet 4.6 / Haiku 4.5 stay current. Claude 3.x is stale-model-id. --fix rewrites 3.x to claude-sonnet-5.

Computer-use safety

computer-unscoped fails closed when computer use / computer_toolset is paired with bare Bash on a jump host.

Out of beta

Agent Skills no longer need the skills-2025-10-02 header. beta-header-stale flags copy that still says they do. plugin-colon-name catches plugin:skill in name:.

Current model IDs

Anything else is stale-model-id. Override leftovers with --allow-model.

claude-opus-5

Opus 5

Deep reasoning, long-horizon agents.

claude-opus-4-8

Opus 4.8

Previous flagship, still current.

claude-sonnet-5

Sonnet 5

Default coding / NetOps agent. --fix target.

claude-sonnet-4-6

Sonnet 4.6

Still current in Claude Code.

claude-haiku-4-5

Haiku 4.5

Fast, cheap classification.

claude-fable-5

Fable 5

Frontier class. Alias: fable.

claude-mythos-5

Mythos 5

Cyberdefense / trusted access. Alias: mythos.

Install

Python 3.8+, no dependencies. Exit non-zero on ERROR so it drops into CI. --fail-on-warn and --min-score 80 for stricter gates.

Install script

curl -fsSL https://raw.githubusercontent.com/gesh75/claude-skill-lint/main/install.sh -o install.sh
bash install.sh
skill-lint ~/.claude/skills

Single file

curl -O https://raw.githubusercontent.com/gesh75/claude-skill-lint/main/skill_lint.py
chmod +x skill_lint.py
./skill_lint.py ~/.claude/skills

GitHub Action

name: skill-lint
on: [push, pull_request]
jobs:
  lint:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v5
      - uses: gesh75/claude-skill-lint@v0.5.0
        with:
          path: skills
          profile: claude-code
          sarif: skill-lint.sarif
      - uses: github/codeql-action/upload-sarif@v3
        if: always()
        with:
          sarif_file: skill-lint.sarif

CLI

skill_lint.py [PATH] [--json] [--sarif FILE] [--profile claude-code|spec|claude-ai]
              [--allow-model ID] [--fail-on-warn] [--fix] [--stdin]
              [--min-score N] [--ignore CODE] [--exclude GLOB]

Profiles: claude-code (default, CC extensions are INFO), spec (strict Agent Skills), claude-ai (reserved names + XML). Score: errors −16, warns −6, infos −1, floor 0.

Checks

89 rules across schema, authoring, safety, disclosure, and style. Full table in the README. New in 0.5:

Code	Level	What it catches
computer-unscoped	WARN	computer use / computer_toolset paired with unscoped Bash
beta-header-stale	INFO	skills-2025-10-02 treated as still required — Agent Skills are out of beta
plugin-colon-name	WARN	name contains a colon; plugin namespace is /plugin:name at invoke time
stale-model-id	WARN	outdated Claude model id (Claude 3.x, pre-4.5 leftovers)
unscoped-bash	WARN	allowed-tools grants bare Bash
disk-wipe / reverse-shell	ERROR	mkfs, dd if=/dev/zero, nc -e, /dev/tcp, pty.spawn

Two cost surfaces

L1 description is loaded every turn (~100 tok). L2 body loads on fire (<5k tok). L3 references/scripts are zero until read. A vague trigger is a permanent tax.

Fail-closed CI

Non-zero exit on ERROR. GitHub Actions annotations and SARIF 2.1.0. Treat a skill like software you install.

Safety second pass

Secrets, reverse shells, disk wipes, curl|bash, insecure TLS, unpinned installs, bundled .env / id_rsa.

NetOps pack

22 production skills. Each file lints clean at ERROR/WARN, has a do-not-use boundary, numbered steps, and a verify loop. Copy into ~/.claude/skills//SKILL.md.

netops-change-guardConfirmed-commit + rollback sdwan-overlay-architectOverlay / DIA / SLA classes sase-zero-trustZTNA / SSE / posture bgp-evpn-closVXLAN EVPN fabric senior-ne-runbookSEV triage, RCA, TAC network-manager-cadenceCAB, capacity, risk netbox-source-of-truthIPAM / DCIM intent multivendor-cli-translateIOS-XE/XR, Junos, EOS, SR Linux firewall-policy-reviewShadow rules, any-any, logging observability-dual-signalAlert only on two signals ipv6-dual-stackGUA, ULA, ND, NAT64 dc-fabric-upgradeISSU / SMU with abort wan-circuit-triageLight levels to ISP RFO campus-nac-dot1x802.1X / MAB / ISE dns-dhcp-ipamAnycast DNS, lease hygiene k8s-cni-networkCNI, NetworkPolicy, egress cloud-hybrid-connectDX / ER / Interconnect / TGW mcp-netops-safetyMCP tools on a jump host intent-nornir-napalmNornir + NAPALM dry-run wifi7-campusWi-Fi 6E/7 RF and MLO mpls-l3vpnPE-CE VRF, RD/RT config-diff-preflightSecrets and blast radius

Source: skills/netops. Lint the pack with python3 skill_lint.py skills/netops --quiet --min-score 90.