harness

chalk-bag

Reusable agent skills for structured software development workflows. Two plugins:

• harness — Documentation lifecycle management with self-improving review agents
• pr — Pull request lifecycle automation

Plugins

harness

A 3-tier documentation system with living execution plans, adversarial code review, conversation mining, and self-improving agent evolution.

Workflow: brainstorm -> plan -> orchestrate -> review -> reflect -> evolve -> complete

Command	Purpose
/harness:init	Initialize 3-tier documentation structure
/harness:brainstorm	Design through collaborative dialogue
/harness:bug	Systematic bug investigation with architecture review
/harness:refactor	Scope incremental refactoring with strangler fig patterns
/harness:plan	Create living execution plans from design docs
/harness:orchestrate	Execute plans with agent teams and micro-reflects
/harness:batch	Execute plans via worktree-isolated parallel batch
/harness:review	Multi-agent code review with adversarial production review
/harness:evolve	Classify learnings, update metrics, propose agent evolution
/harness:reflect	Full reflection, conversation mining, retrospective
/harness:complete	Archive plan, prune check, and create PR
/harness:prune	Audit docs for staleness, broken links, bloat

Agents: harness-pruner, learnings-reviewer, harness-evolver

Skills: strangler-fig (incremental refactoring patterns)

pr

Pull request lifecycle management with multi-perspective automated review.

Command	Purpose
/pr:author	Create PRs with quality gates
/pr:automate	Full automated lifecycle: author -> review -> resolve -> merge
/pr:review	Multi-agent PR review (6 specialized agents)
/pr:resolve	Analyze and address PR review comments
/pr:update	Sync PR description with current changes

Installation

Claude Code

Install as a marketplace. From any project directory:

claude /plugins add https://github.com/donovan-yohan/chalk-bag

This registers both plugins. You can also install individual plugins:

# Install just the harness plugin
claude /plugins add https://github.com/donovan-yohan/chalk-bag/plugins/harness

# Install just the pr plugin
claude /plugins add https://github.com/donovan-yohan/chalk-bag/plugins/pr

Hermes

Hermes can load these skills through its plugin and profile system. See docs/hermes.md for integration instructions.

Other Agents

The command and skill files are standard markdown with YAML frontmatter. Any agent that can load markdown-based instructions can use these skills directly. The key integration points:

1. Commands (plugins/*/commands/*.md) — Procedural workflows triggered by explicit user commands
2. Agents (plugins/*/agents/*.md) — System prompts for specialized background agents
3. Skills (plugins/*/skills/*/SKILL.md) — Pattern-triggered capabilities with frontmatter descriptions
4. Scripts (plugins/harness/scripts/*.sh) — Shell scripts for persistence and metrics
5. References (plugins/harness/references/*.md) — Shared reference documents loaded by commands

Dependencies

The harness plugin works standalone. Some commands optionally integrate with:

• superpowers — Used by brainstorm, plan, orchestrate, review, and complete commands for their core methodologies (brainstorming, writing-plans, subagent-driven-development, etc.)
• pr-review-toolkit — Used by the review command for specialized review agents (code-reviewer, silent-failure-hunter, type-design-analyzer)

The pr plugin requires pr-review-toolkit for the review and automate commands.

Documentation Tiers

The harness plugin manages a 3-tier documentation system:

1. CLAUDE.md (60-120 lines) — Map with Documentation Map table
2. docs/*.md — Domain summaries (ARCHITECTURE, DESIGN, PLANS, QUALITY)
3. docs/design-docs/, docs/exec-plans/ — Deep docs, versioned plans

Self-Improvement

The harness plugin tracks review effectiveness across sessions:

• Metrics — Review agent accuracy, plan prediction quality, learning efficacy
• Evolution — Automatic agent definition updates based on review escapes
• Adversarial review — Context-isolated production failure analysis using claude -p
• Learnings — Persistent project knowledge captured and enforced across sessions

License

MIT