wos

Experimental — This project is under active development. APIs, skill

interfaces, and document conventions may change without notice.

A Claude Code plugin marketplace for building and maintaining structured project context, with AI-assisted research, source verification, and quality validation.

For Claude Code users who want reusable tooling for skill authoring, project documentation, task planning, and structured reasoning — packaged as four independently installable plugins.

Prerequisites

• •Python 3.9+
• •No runtime Python dependencies (stdlib only)
• •gh CLI (for work:finish-work skill)

Plugins

Four self-contained plugins, each independently installable:

Installation

Knowledge — build and maintain project context documents:

claude plugin install bcbeidel/toolkit --plugin wiki

Build tooling — create, audit, and maintain Claude Code skills and rules:

claude plugin install bcbeidel/toolkit --plugin build

Full suite — complete toolkit for AI-assisted project work:

claude plugin install bcbeidel/toolkit --plugin wiki
claude plugin install bcbeidel/toolkit --plugin build
claude plugin install bcbeidel/toolkit --plugin work
claude plugin install bcbeidel/toolkit --plugin consider

Usage

Once installed, invoke a plugin's skills via the /plugin:skill-name syntax in Claude Code. For example, after installing the wiki plugin:

/wiki:setup

Expected: Claude Code initializes project context directories and prompts for any missing metadata.

See each plugin's skill directory for the full skill catalogue (plugins/<plugin>/skills/).

Security scanning

Every PR that touches plugins/, policy/, or the scanner pipeline runs the Cisco AI Skill Scanner plus a Claude evaluator and posts a single Skill Security Audit Report comment on the PR. The gate blocks merge on overall_severity=high unless the PR carries the security-override label.

The scanner runs in two coupled workflows:

• •.github/workflows/skill-audit-scan.yml — pull_request, contents: read

only, runs against the PR head. Produces a scan-<plugin> artifact.

• •.github/workflows/skill-audit-report.yml — workflow_run, checks out

trusted main, posts the comment, enforces the gate.

Detection rules (taxonomy, escalation signals, prohibited extensions) live in `policy/skill-scan-policy.yml`. The SHA256 of that file is recorded as policy_fingerprint on every scan.

Run locally before pushing:

export ANTHROPIC_API_KEY=<key>
make scan PLUGIN=build      # one plugin
make scan-all               # every plugin
make scan-clean             # remove scan-output/

make scan mirrors the CI workflow: same policy file, same hash-locked deps in .github/scripts/requirements.lock, same --fail-on-severity high threshold. A clean local scan should match what CI produces on the same SHA.

Composition with /build:check-skill: that skill audits authorship quality (structure, completeness, prompt clarity); skill-audit-* audits adversarial content (prompt injection, exfiltration, supply-chain risk). They are independent, complementary signals.

Local Development

pip install -e plugins/wiki -e ".[dev]"

python -m pytest plugins/wiki/tests/ -v

ruff check plugins/

Project Structure

plugins/
  build/       # Skill authoring and auditing
  wiki/        # Project context: setup, research, ingest, lint
    src/wiki/  # Python package
    scripts/   # Shared CLI scripts
    skills/    # Skill definitions
    tests/     # Unit tests
  work/        # Task scoping, planning, execution
  consider/    # Structured mental models
.context/      # Domain knowledge and conventions
.plans/        # Implementation plans
.research/     # Research investigations
.prompts/      # Saved and refined prompts
.claude-plugin/
  marketplace.json

Contributing

See CONTRIBUTING.md.

License

MIT — see LICENSE.