Claude Night Market

A plugin marketplace for Claude Code, Anthropic's agentic coding tool.

Night Market extends Claude Code with 23 plugins covering git workflows, code review, spec-driven development, architecture selection, codebase visualization, autonomous agents, multi-LLM delegation, ML-enhanced scoring, and multi-source research. 188 skills, 131 slash commands, and 56 agents. Each plugin installs independently.

Highlights

• TDD by default. imbue enforces failing-test-first via PreToolUse hooks; CONSTITUTION.md encodes rules that override any conflicting skill.
• Layered architecture. Four internal layers (Foundation, Utility, Domain, Meta) prevent dependency cycles. Install only what you need.
• Multi-LLM delegation. conjure routes tasks to Gemini and Qwen with cheapest-capable model selection.
• AI-slop guards. scribe:slop-detector runs four layers of checks (P0 critical patterns, document economy, sentence-level slop, evidence-backed claims) before docs ship.
• Cross-session state. Task lists, sessions, and decisions persist via CLAUDE_CODE_TASK_LIST_ID and GitHub Discussions.

Contents

• Quick Start
• Trust and Safety
• Architecture (plugin catalog, layer model)
• Common Workflows
• Requirements
• What's New
• Plugin Development
• Documentation
• Stewardship · Contributing · Acknowledgements · License

Quick Start

Requires Claude Code 2.1.16+ and Python 3.9+ for hooks. See Requirements for details.

# Add the marketplace
/plugin marketplace add athola/claude-night-market

# Install plugins you need
/plugin install sanctum@claude-night-market    # Git workflows
/plugin install pensive@claude-night-market    # Code review
/plugin install spec-kit@claude-night-market   # Spec-driven dev

# Use them
/prepare-pr                                    # Prepare a pull request
/full-review                                   # Run code review

Alternative: Install via npx with npx skills add athola/claude-night-market (installs all plugins at once).

After installation, run claude --init for one-time setup.

Note: If the Skill tool is unavailable, read skill files directly at plugins/{plugin}/skills/{skill-name}/SKILL.md.

opkg (OpenPackage)

# Install specific plugins
opkg i gh@athola/claude-night-market --plugins sanctum
opkg i gh@athola/claude-night-market --plugins pensive,conserve

# Plugins that depend on shared runtime skills (e.g. attune, conjure)
# automatically pull packages/core as a dependency

See the Installation Guide for detailed setup options.

Trust and Safety

⚠️ Plugins run inside your Claude Code session and can read or edit your repo, run shell commands, and call external services. Review any plugin before installing it.

Night Market ships three lines of defense, but none replace your own review:

• TDD gates (imbue): a PreToolUse hook blocks implementation writes that lack a corresponding failing test.
• Destructive-command blockers (conserve): auto-approve safe commands while halting rm -rf, git push --force, git reset --hard, and similar.
• Production-environment guard (hookify:destructive-command-guard): warns when a destructive command targets a production-shaped path or env var (prod, PROD_*, live). Defaults to warn, not block, so legitimate ops proceed with friction, not refusal.
• Additive-bias audits (leyline:additive-bias-defense): every diff is checked for unjustified additions before commit.

See STEWARDSHIP.md for the maintenance contract and CONSTITUTION.md for the immutable rules that override any conflicting skill or hook.

Architecture

23 internal plugins in four layers, plus external plugins from the superpowers-marketplace. Arrows show dependency direction (A --> B means A depends on B). Dashed arrows mark optional complements.

Source: assets/architecture.d2; regenerate with d2 assets/architecture.d2 assets/architecture-light.svg

Plugin Catalog

| Plugin | Layer | Description | Skills | Cmds | |--------|-------|-------------|:------:|:----:| | abstract | Meta | Skill authoring, hook development, evaluation frameworks, escalation governance, hook scope guidance | 15 | 18 | | leyline | Foundation | Auth flows (GitHub/GitLab/AWS), quota management, error patterns, markdown formatting, Discussions retrieval, damage-control, stewardship, trust verification, injection detection, deferred-capture contracts, additive-bias defense | 22 | 3 | | sanctum | Foundation | Git workflows, commit messages, PR prep, docs updates, version management, sessions, deferred-item capture | 18 | 20 | | imbue | Foundation | TDD enforcement, proof-of-work validation, scope guarding, additive-bias auditing, rigorous reasoning, vow enforcement | 13 | 5 | | conserve | Utility | Context optimization, bloat detection, context mapping, CPU/GPU monitoring, token conservation, filter-first log debugging | 13 | 5 | | conjure | Utility | Delegation framework for routing tasks to external LLMs (Gemini, Qwen) with cheapest-capable model selection | 4 | 0 | | hookify | Utility | Behavioral rules engine with markdown configuration, hook-to-rule conversion, and a security rule catalog (production-guard, destructive-command-guard) | 2 | 6 | | egregore | Utility | Autonomous agent orchestrator with parallel worktrees, agent specialization, cross-item learning, and crash recovery | 4 | 5 | | herald | Utility | Shared notification library: GitHub issue alerts, webhook support (Slack, Discord, generic) | 0 | 0 | | oracle | Utility | ONNX Runtime inference daemon for ML-enhanced plugin capabilities over localhost HTTP | 1 | 1 | | pensive | Domain | Code review, architecture review, bug hunting, blast radius analysis, performance hotspots, Makefile audits, NASA Power of 10 | 14 | 13 | | attune | Domain | Project lifecycle: brainstorm, specify, plan, interactive plan review, initialize, execute, war-room, dorodango polishing | 13 | 10 | | spec-kit | Domain | Spec-driven development: specifications, task generation, implementation | 3 | 10 | | parseltongue | Domain | Python: testing, performance, async patterns, packaging | 4 | 3 | | minister | Domain | GitHub issue management, label taxonomy, initiative tracking, DORA delivery metrics | 3 | 3 | | memory-palace | Domain | Spatial knowledge organization, digital garden curation, PR review capture | 7 | 5 | | archetypes | Domain | Architecture paradigm selection (hexagonal, CQRS, microservices, etc.) | 14 | 0 | | gauntlet | Domain | Codebase learning through knowledge extraction, challenges, code knowledge graph, and spaced repetition | 7 | 6 | | phantom | Domain | Computer use: screenshot capture, mouse/keyboard control, autonomous desktop agent | 1 | 1 | | scribe | Domain | Documentation, AI slop detection, SICO voice extraction, style transfer, session replay | 11 | 9 | | scry | Domain | Terminal recordings (VHS), browser recordings (Playwright), GIF processing | 4 | 2 | | tome | Domain | Multi-source research: code archaeology, community discourse, academic literature, TRIZ analysis | 7 | 4 | | cartograph | Domain | Codebase visualization: architecture, data flow, dependency, call chains, community detection, class diagrams via Mermaid | 7 | 1 |

Full inventory: Capabilities Reference.

How the Layers Work

Governance. imbue enforces TDD via a PreToolUse hook that verifies test files before allowing implementation writes. Quality gates halt execution when tests fail.

Security. leyline manages OAuth flows with local token caching. conserve auto-approves safe commands while blocking destructive operations. sanctum isolates named sessions, and agents can run in worktree isolation for parallel execution.

Orchestration. egregore manages autonomous agent lifecycles with parallel worktree execution, agent specialization, cross-item learning, and crash recovery via watchdog monitoring.

Maintenance. /update-ci reconciles pre-commit hooks and GitHub Actions with code changes. abstract tracks skill stability and auto-triggers improvement agents when degradation is detected.

Cross-session state. attune, spec-kit, and sanctum persist state across sessions via CLAUDE_CODE_TASK_LIST_ID. GitHub Discussions serve as a second persistence layer for decisions, war-room deliberat