coreteam-claude-base — reusable Claude Code plugin marketplace

A project-agnostic Claude Code plugin marketplace. Ships generic skills, slash commands, sub-agents, and hooks plus an example layout you can fork into your own org's marketplace.

Replace <owner> in the badge URL above with the GitHub org once this repo's home is final.

⚠️ This codebase lives in TWO repositories — they MUST stay in sync

This project is published from both:

• PerfectoMobileDev/claude-base — https://github.com/PerfectoMobileDev/claude-base
• Blazemeter/claude-base — https://github.com/Blazemeter/claude-base

Every commit on every branch must appear on both. An auto-mirror GitHub Action propagates pushes from one side to the other; a strict CI check (Verify sibling sync) blocks merges to main when the two diverge. See LINKED_REPOS.md for the contract, the recovery procedure, and the one-time PAT / branch-protection setup.

Quick install (for users)

Inside Claude Code:

/plugin marketplace add <owner>/<repo>
/plugin install base-tools@coreteam-claude-base

(Replace <owner>/<repo> with the GitHub path where this repo lives.)

That's it — Claude will auto-load all skills, register all slash commands, expose all sub-agents, and activate plugin hooks. Updates land automatically when you run /plugin update.

To verify:

/plugin list                          # confirms base-tools is installed
/base-tools:example-command           # try the example slash command

To update and reload the plugin:

/plugin marketplace update coreteam-claude-base
/reload-plugins

What's inside

coreteam-claude-base/
├── .claude-plugin/
│   └── marketplace.json              # Catalog — lists every plugin in this repo
├── plugins/
│   └── base-tools/                   # The (currently single) bundled plugin
│       ├── .claude-plugin/
│       │   └── plugin.json           # Plugin manifest (name, version, license)
│       ├── skills/                   # Auto-loaded by Claude based on description
│       │   ├── example-skill/
│       │   │   └── SKILL.md          # Template — copy to start a new skill
│       │   └── summarize-pr/         # Real working skill — reference implementation
│       │       ├── SKILL.md
│       │       └── tests/
│       │           └── cases.yaml    # Behavioral test cases
│       ├── commands/                 # Slash commands (/base-tools:<name>)
│       │   └── example-command.md    # Template — copy to start a new command
│       ├── agents/                   # Sub-agents (Agent tool / /agents menu)
│       │   └── example-agent.md      # Template — copy to start a new agent
│       ├── hooks/                    # Hook scripts referenced from hooks.json
│       │   ├── README.md
│       │   ├── audit-file-changes.sh
│       │   ├── block-secrets-on-write.sh
│       │   ├── block-ec2-cloud-mutations.sh    # Safety: compute is read-only
│       │   ├── block-datastore-mutations.sh    # Safety: DynamoDB/S3/Redis/Mongo read-only
│       │   ├── block-shared-branch-git.sh      # Safety: protect master/shared branches
│       │   ├── block-credentials-in-commit.sh  # Safety: no key files / secrets in commits
│       │   ├── block-test-skips.sh             # Safety: don't skip failing tests
│       │   ├── block-hook-bypass.sh            # Safety: don't bypass the hooks
│       │   └── log-skill-activation.py
│       └── hooks.json                # Plugin-level event hooks (Pre/Post ToolUse, Skill, etc.)
├── policy/
│   └── allowed-tools.yaml            # Tool allowlist + denylist enforced by validator
├── scripts/
│   ├── validate.py                   # Python validator (schema, standards, policy, secrets)
│   ├── behavioral_runner.py          # Schema-lint + opt-in Anthropic-API skill tests
│   ├── aggregate_telemetry.py        # Summarize skill-activation telemetry
│   ├── requirements.txt              # pip dep: pyyaml
│   ├── requirements-eval.txt         # Optional: anthropic SDK (for API runs)
│   ├── README.md
│   └── tests/
│       └── test_validate.py          # Regression tests (stdlib only)
├── .github/
│   ├── CODEOWNERS
│   ├── PULL_REQUEST_TEMPLATE.md
│   └── workflows/plugin-validate.yml # CI: structural / SAST / secret-scan / claude-cli
├── CONTRIBUTING.md
├── LICENSE
└── README.md

Adding new artifacts

A new skill

cp -r plugins/base-tools/skills/example-skill plugins/base-tools/skills/<your-skill>
$EDITOR plugins/base-tools/skills/<your-skill>/SKILL.md

Edit the YAML frontmatter:

• name — kebab-case ID (must match directory name).
• description — Claude reads this to decide when to auto-load the skill. Be specific about triggers ("Use when the user asks to X").
• allowed-tools — optional space-separated list of tools the skill can use without prompting.

Once committed, users get it on the next /plugin update.

A new slash command

cp plugins/base-tools/commands/example-command.md plugins/base-tools/commands/<your-command>.md

Slash commands are flat .md files (no subdirectory). After install they're invokable as /base-tools:<your-command>. They support $ARGUMENTS, $1, $2, and dynamic context injection with !`shell-cmd`.

A new sub-agent

cp plugins/base-tools/agents/example-agent.md plugins/base-tools/agents/<your-agent>.md

Sub-agents are flat .md files with name, description, model, and tools frontmatter. They show up under /agents and can be called from the main agent via the Agent tool with subagent_type: "<your-agent>".

A new hook

Hook scripts live under plugins/base-tools/hooks/ (one script per file). The wiring is in plugins/base-tools/hooks.json, which maps events + matchers to script paths.

Workflow:

# 1. Drop your script in the hooks folder. Use a shebang.
$EDITOR plugins/base-tools/hooks/<your-hook>.sh

# 2. Register it in hooks.json under the right event:
#    "<EventName>": [{ "matcher": "<regex>", "hooks": [{ "type": "command",
#      "command": "${CLAUDE_PLUGIN_ROOT}/hooks/<your-hook>.sh" }] }]

# 3. (Linux/macOS) make it executable
chmod +x plugins/base-tools/hooks/<your-hook>.sh

Supported events:

• PreToolUse, PostToolUse (with matcher regex on tool name) — non-zero exit on Pre* blocks the tool call.
• PreCommand, PostCommand
• PreFile, PostFile
• PreSkill, PostSkill
• PostSubagentSpawn

Environment provided to hook scripts:

• ${CLAUDE_PLUGIN_ROOT} — install path of this plugin (use for command: references in hooks.json and for sourcing helpers).
• ${CLAUDE_PLUGIN_DATA} — per-user persistent data dir (use for logs, audit state — survives plugin upgrades).
• ${CLAUDE_SESSION_ID} — current session.
• stdin — the tool/event payload as JSON. Parse with jq or json.load(sys.stdin).

See plugins/base-tools/hooks/README.md for full conventions and the bundled example scripts:

• audit-file-changes.sh — PostToolUse logger for Write/Edit calls.
• block-secrets-on-write.sh — PreToolUse guard that blocks writes containing obvious secret patterns.
• log-skill-activation.py — PreSkill telemetry collector.

Safety guardrails (block destructive actions Claude shouldn't take autonomously — see STANDARDS.md → "Safety guardrails"):

• block-ec2-cloud-mutations.sh — no AWS EC2 / GCP Compute instance create/start/stop/terminate/delete (read-only allowed).
• block-datastore-mutations.sh — no destructive DynamoDB / S3 / Redis / Mongo CLI ops (reads/downloads/dumps allowed).
• block-shared-branch-git.sh — no push to master/main, no force push / rebase / reset --hard on shared branches.
• block-credentials-in-commit.sh — no committing key/credential files or staged secrets.
• block-test-skips.sh — no disabling/skipping failing tests to green CI.
• block-hook-bypass.sh — no --no-verify, core.hooksPath override, or inlined override prefix to bypass the hooks above.

A human-exported CLAUDE_SAFETY_OVERRIDE=1 bypasses these for a session (logged to ${CLAUDE_PLUGIN_DATA}/base-tools/safety-override.log).

Adding a second plugin

This marketplace supports many plugins. To add one:

1. mkdir -p plugins/<new-plugin>/.claude-plugin/
2. Create plugins/<new-plugin>/.claude-plugin/plugin.json (copy base-tools/.claude-plugin/plugin.json and rename).
3. Add an entry to the top-level .claude-plugin/marketplace.json plugins array.
4. Drop skills/, commands/, agents/, hooks/, hooks.json under the new plugin directory.

Users then install with /plugin install <new-plugin>@coreteam-claude-base.

Validating before pushing

Five independent CI jobs gate every merge. Each runs on its own runner — a failure in one diagnoses without blocking insight into the others.

Layer 1 — structural / standards / tool policy / smoke secrets (Python)

pip install -r scripts/requirements.txt   # one-time, single dep (pyyaml)
python scripts/validate.py                # report
python scripts/validate.py --strict       # treat warnings as errors

scripts/validate.py covers:

• Schema for marketplace.json, every plugin.json, every SKILL.md / command / agent frontmatter, and hooks.json.
• Standards: kebab-case names, name ↔ directory parity, SKILL.md length budget, agent file-stem ↔ name parity, no human-centric files inside skill dirs.
• Hook wiring: every referenced script exists; each script has a shebang; every event name is known.
• Tool policy: every allowed-tools (skills/commands) and tools (agents) entry is checked against policy/allowed-tools.yaml. Unscoped Bash is forbidden by default. MCP tools (mcp__*) are exempt.
• Smoke secret scan: high-signal patterns (AWS keys, private key headers, GH / Slack / GitLab / Google tokens) so contributors catch obvious leaks before pushing.

See scripts/README.md for the full check list, the tool policy, and instructions for adding project-specific rules.

Layer 2 — SAST on hook scripts (shellcheck, ruff)

Hook scripts run unsupervised on every developer's machine — they're the highest-leverage surface. CI lints them:

• shellcheck plugins/**/hooks/**/*.sh — catches unquoted vars, missing error handling, dangerous patterns.
• ruff check plugins/**/hooks/**/*.py — catches bare except, eval-of-user-input, broken imports, lint.

Replicate locally:

shellcheck plugins/*/hooks/*.sh
pip install ruff && ruff check plugins/*/hooks/*.py

Layer 3 — comprehensive secret scan (gitleaks)

CI runs gitleaks/gitleaks-action with fetch-depth: 0 so it scans the whole history, not just the PR diff. To replicate locally:

gitleaks detect --source . --redact

Layer 4 — Claude Code's own validator

# Inside Claude Code:
/plugin validate .

# Or from the terminal (requires the Claude CLI):
claude plugin validate .

This is the canonical schema check that mirrors what Claude Code does at plugin-install time.

Layer 5 — Behavioral tests (opt-in)

Each skill can ship tests/cases.yaml with prompts + assertions. scripts/behavioral_runner.py lints them every PR and — if the ANTHROPIC_API_KEY secret is configured — runs them against the Anthropic API:

python scripts/behavioral_runner.py --lint                  # always works
ANTHROPIC_API_KEY=... pip install -r scripts/requirements-eval.txt
python scripts/behavioral_runner.py                          # API mode

See plugins/base-tools/skills/summarize-pr/tests/cases.yaml for a worked example.

Supply-chain hardening

Every action used in .github/workflows/plugin-validate.yml is pinned by full commit SHA (with # vN comment for readability) — a compromised tag can't escalate into CI. When bumping versions:

gh api repos/<owner>/<repo>/commits/v<N> --jq .sha

Replace the SHA in the workflow and update the trailing comment. Never use a floating tag (@v4, @main).

Release process

• Default: leave "version" out of plugin.json — Claude Code uses the git commit SHA so every merge to main is automatically the latest release.
• Explicit versioning: set "version": "x.y.z" and bump it for every user-visible release. Users won't get updates until the string changes.
• Channels: maintain two marketplace entries (e.g. stable and latest refs) if you need staged rollouts.

Why this format

Claude Code natively discovers plugins published as marketplaces, so users get auto-loading skills, namespaced slash commands, an /agents menu, plugin hooks, and one-command updates. See the official Claude Code plugin docs for the full spec.

License

See LICENSE (if present) at repo root. The base-tools plugin manifest is set to UNLICENSED by default — change it to MIT, Apache-2.0, or whichever SPDX identifier applies before publishing externally.

{
  "name": "coreteam-claude-base",
  "description": "Reusable Claude Code plugin marketplace — generic skills, slash commands, sub-agents, and hooks. Use as a starting point or fork into project-specific marketplaces.",
  "owner": {
    "name": "coreteam-claude-base maintainers"
  },
  "plugins": [
    {
      "name": "base-tools",
      "source": "./plugins/base-tools",
      "description": "Base plugin with generic skill, slash-command, sub-agent, and hook templates plus working hook examples (audit logger, secret-on-write guard, skill-activation telemetry).",
      "version": "0.3.0",
      "keywords": [
        "base",
        "templates",
        "skills",
        "commands",
        "agents",
        "hooks",
        "starter"
      ],
      "category": "templates"
    },
    {
      "name": "mend",
      "source": "./plugins/mend",
      "description": "Mend (WhiteSource) SCA knowledge — REST v1.3 auth, project-token lookup, alert fetch, CVE severity. Reusable read layer.",
      "version": "0.1.0",
      "keywords": [
        "mend",
        "whitesource",
        "sca",
        "cve",
        "security"
      ],
      "category": "development"
    },
    {
      "name": "dep-remediation",
      "source": "./plugins/dep-remediation",
      "description": "Fix vulnerable dependencies correctly — golden rule (direct/BOM not transitive), advisory cross-check, defer breaking majors, per-stack recipes (composer/gradle/maven/pip/npm) + local build+test.",
      "version": "0.3.2",
      "keywords": [
        "dependencies",
        "remediation",
        "composer",
        "gradle",
        "maven",
        "pip",
        "npm",
        "bom"
      ],
      "category": "development"
    },
    {
      "name": "jenkins",
      "source": "./plugins/jenkins",
      "description": "Blazemeter blazect Jenkins gate — trigger a multibranch build with PUSH_TO_GCR + PERFORM_WHITESOURCE_SCAN, gate on the specific triggered build (not lastBuild), fix-forward, never gate on API_TESTS-CI; plus deploy a green branch to denv (USERNAME=JENKINS_USER) via the shared 'Deploy environment' job and gate on it.",
      "version": "0.1.6",
      "keywords": [
        "jenkins",
        "ci",
        "blazect",
        "gcr",
        "build-gate",
        "denv-deploy"
      ],
      "category": "development"
    },
    {
      "name": "github",
      "source": "./plugins/github",
      "description": "Blazemeter GitHub branch/PR conventions — dated Docker-safe branch, commit+push via SSO gh, PR into the integration branch, tag PR title with the ticket id (in the title at creation when the ticket already exists).",
      "version": "0.2.0",
      "keywords": [
        "github",
        "pull-request",
        "branch",
        "gh"
      ],
      "category": "development"
    },
    {
      "name": "jira",
      "source": "./plugins/jira",
      "description": "Create/transition Blazemeter MOB Jira tickets — board 5348, Task, Product/Scrum-Team/Sprint fields, assignee = repo owner, status In Progress, caller-supplied labels (e.g. mend_orch), optional parent issue (parent_key in jira config), bounded 429 backoff. Ticket summary/name is caller-supplied. Can be created before a PR exists, then updated with the PR link.",
      "version": "0.3.2",
      "keywords": [
        "jira",
        "mob",
        "atlassian",
        "ticket"
      ],
      "category": "development"
    },
    {
      "name": "mend-blz",
      "source": "./plugins/mend-blz",
      "description": "End-to-end Mend remediation recipe — composes mend + dep-remediation + jenkins + github + jira into one loop (alerts → fix → Jenkins green → [crane only] a.blazemeter.com branch → denv deploy + fix-related API tests → Confluence report → Jira created first → PR titled with the ticket id → Jira description updated with the PR link → [crane only] mend-finalize follow-up message), registry-driven via services.json.",
      "version": "0.7.2",
      "keywords": [
        "mend",
        "whitesource",
        "security",
        "cve",
        "vulnerabilities",
        "blazemeter",
        "recipe"
      ],
      "category": "development"
    },
    {
      "name": "mend-finalize",
      "source": "./plugins/mend-finalize",
      "description": "Phase 2 of the crane Mend flow — after the crane PR merges to the integration branch and the CI build completes, fetches the real semantic version, updates HarborVersionsSettings.php in the a.blazemeter.com branch from latest-<branch> to the real version, opens the a.blazemeter.com PR into develop, and updates the MOB ticket. Run after merging the crane PR opened by mend-blz.",
      "version": "0.1.3",
      "keywords": [
        "mend",
        "crane",
        "harbor",
        "version",
        "blazemeter",
        "finalize"
      ],
      "category": "development"
    },
    {
      "name": "opensearch",
      "source": "./plugins/opensearch",
      "description": "Search BlazeMeter logs across denv/bzdev/ci/staging/prod — bundles a read-only opensearch-reader MCP server plus a skill that builds safe Lucene queries and condenses results.",
      "version": "0.1.0",
      "keywords": [
        "opensearch",
        "logs",
        "observability",
        "lucene",
        "bzdev",
        "denv",
        "search"
      ],
      "category": "observability"
    },
    {
      "name": "master-investigation",
      "source": "./plugins/master-investigation",
      "description": "Structured playbook for investigating a misbehaving BlazeMeter master/test/session (stuck, wrong metric, silent failure, intermittent, customer-reported) across denv/bzdev/ci/staging/prod — composes OpenSearch log search, Mongo (main + bigdata) state, GCS session artifacts, and source-code verification across taurus/taurus-cloud/a.blazemeter.com/bzm-crane.",
      "version": "0.2.1",
      "keywords": [
        "blazemeter",
        "master",
        "session",
        "taurus",
        "investigation",
        "incident",
        "debugging"
      ],
      "category": "observability"
    },
    {
      "name": "pr-review",
      "source": "./plugins/pr-review",
      "description": "Address and incorporate GitHub PR review comments for any repo — fetch, validate against the repo's own conventions, incorporate minimally, auto-detect and run tests/lint, one-pass code-review sub-agent, commit, reply to every thread, optional polling.",
      "version": "0.1.0",
      "keywords": [
        "pr",
        "pull-request",
        "code-review",
        "github",
        "review-comments"
      ],
      "category": "development"
    },
    {
      "name": "ci-triage",
      "source": "./plugins/ci-triage",
      "description": "Triage backend API_TESTS-CI failure noise from the BlazeMeter Slack monitoring channels: find persistent vs. flaky failures, identify the likely culprit commit, and check for an already-open PR before ever drafting a notification.",
      "version": "0.1.0",
      "keywords": [
        "ci",
        "jenkins",
        "slack",
        "triage",
        "blazemeter",
        "api-tests"
      ],
      "category": "development"
    }
  ]
}