Warden — Development skill for Claude Code
Runtime verification layer for Claude Code skills: observe every tool call, enforce declared capabilities, record replayable trajectories.
How to install Warden
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open Hans-wan/warden and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Warden does
Runtime verification layer for Claude Code skills: observe every tool call, enforce declared capabilities, record replayable trajectories.
Alternatives in Development
- Dive — Dive is an open-source MCP Host Desktop Application that seamlessly integrates with any LLMs supporting functi 1.8k ★
- Claude Engineer — A version of command line interface(CLI) that leverages the power of Claude Opus(3-4.7) 954 ★
- WP Abilities API — WordPress capabilities and permissions API 902 ★
README
[](LICENSE) [](https://en.cppreference.com/w/cpp/20) [](https://code.claude.com/docs/en/plugins) [](#doctor) [](CONTRIBUTING.md)
**English** · [简体中文](README.zh-CN.md) · [Design doc](docs/superpowers/specs/2026-09-22-warden-design.md)
**⚠️ warden is not a security boundary.** The best-effort regexes can be bypassed via `bash -c "$(...)"`, base64, `python -c`, and other indirect execution. v1 exists to **help honest skills show what they actually did** — not to stop malicious ones. Adversarial defense lands in v2 (sandboxing).
Why warden
Installing a skill on your agent is a stack of verbal promises: "I only read files", "I only run git status". warden turns those promises into **verifiable runtime facts**:
| Observe | every tool call is recorded, with sanitized arguments |
| Judge | calls that exceed the skill's declared capabilities are blocked or surfaced per policy |
| Replay | trajectories can be replayed and diffed; skill drift shows up in CI |
Quick Start
Install
Requires CMake ≥ 3.20, Ninja, a C++20 compiler.
cmake -B build -G Ninja && cmake --build build
install -D build/warden /bin/warden
warden ships as a Claude Code plugin (`.claude-plugin/plugin.json` + `hooks/hooks.json`) — hooks mount automatically, **zero user configuration**. The host expands `${CLAUDE
Related Skills
Retrospective Harness
Trigger deterministic retrospective analysis of recent agent trajectories, extract invariant friction signatur
Agent Contract Nexus
Agentic Governance 2026: Zero-Runtime Skills, Rules & Replayable State Contracts
App Capabilities
/app-capabilities — enforce role boundaries
Skill Feedback
Rate a skill's performance after use (self-improving skills OBSERVE layer — user feedback)
Session Verify
Dump every kit-loaded resource — auto-memory dir, MEMORY.md disk-vs-runtime parity, working-folder files, work
Greenfield Run With Feedback Loop
Instrumented AEF greenfield install — dual mandate (install AND falsify), local-first findings harness with a
Related Agents
Beads Warden
Guard the beads execution record: enforce the write-flush-verify discipline that defeats the bd rapid-write ra
Goal Contract Auditor
Audits the trust contract against the code. Invoke after changing anything the engine prints, adding a record
Autopilot Orchestrator
Full RTL design pipeline orchestrator. Manages 6-phase flow with dual-layer phase gates, parallel agent execut