Setup Harness Skill

Copies the Claude Code harness into project directories so it becomes part of the Git repository.

Trigger Patterns

• User says "setup harness"
• User provides a path to claude-code-harness
• User says "copy the harness", "install harness configuration"
• User asks how to use the harness in a new project

What This Skill Does

Deploys harness files from the source to target project(s) using deploy-harness.sh:

1. Syncs .claude/ directory (excluding runtime files) via rsync
2. Creates runtime directories with .gitkeep files
3. Updates .gitignore for runtime exclusions
4. Installs git hooks (doc-gardener pre-push)
5. Optionally copies .mcp.json

Why Copy Instead of Symlink?

• Version Control: .claude/ becomes part of your repo
• Self-Contained: No external dependencies
• CI/CD Ready: Works in pipelines without the source harness
• Project-Specific: Customize without affecting other projects

Deployment Script

All deployment logic is implemented in deploy-harness.sh.

The script lives at .claude/skills/setup-harness/deploy-harness.sh and handles the full deployment lifecycle: source validation, rsync with exclusions, runtime directory creation, .gitignore updates, git hook installation, and verification.

Usage

Deploy to all configured targets:

.claude/skills/setup-harness/deploy-harness.sh

Deploy to a specific path:

.claude/skills/setup-harness/deploy-harness.sh --target ~/Documents/Windsurf/my-project

Deploy to a named target:

.claude/skills/setup-harness/deploy-harness.sh --name my-project

Preview without changes:

.claude/skills/setup-harness/deploy-harness.sh --dry-run

List configured targets:

.claude/skills/setup-harness/deploy-harness.sh --list

Include .mcp.json:

.claude/skills/setup-harness/deploy-harness.sh --target ~/proj --include-mcp

Targets Configuration

Deployment targets are configured in targets.json (same directory as this skill). Edit targets.json to add or remove targets. The ~ in paths is expanded at runtime.

Interactive Workflow (When Claude Runs This Skill)

When triggered by a user, Claude should:

Step 1: Determine Target

If the user specified a path, use it directly. Otherwise, ask:

Question: "Where do you want to set up the Claude Code harness?"
Header: "Target Dir"
Options:
1. "All configured targets (Recommended)" - Deploy to all targets in targets.json
2. "Specify path" - Provide a custom directory path

multiSelect: false

Step 2: Handle .mcp.json

Question: "How do you want to handle .mcp.json?"
Header: "MCP Config"
Options:
1. "Skip it (Recommended)" - Don't copy .mcp.json (API keys differ per project)
2. "Copy it" - Copy .mcp.json to target (remember to update API keys)

multiSelect: false

Step 3: Run the Script

Based on user choices, construct and run the appropriate command:

# Example: Deploy to all targets without .mcp.json
.claude/skills/setup-harness/deploy-harness.sh

# Example: Deploy to specific path with .mcp.json
.claude/skills/setup-harness/deploy-harness.sh --target /path/to/project --include-mcp

Step 4: Report Results

The script outputs verification results. Summarize for the user and remind them to:

1. Review .claude/CLAUDE.md (harness docs — updated each deploy)
2. Review .mcp.json API keys (if copied)
3. Commit the .claude/ directory to git

What the Script Does (Reference)

The following steps are all handled by deploy-harness.sh. They are documented here for reference only — Claude should NOT execute these manually.

Source Validation

• Checks harness source exists with settings.json and skills/
• Warns about stale state/progress files in source

rsync with Exclusions

Copies .claude/ while excluding runtime artifacts:

• /state/* — Runtime state files
• /completion-state/ — Session completion tracking
• /progress/* — Session progress files
• /worker-assignments/* — Worker task assignments
• /logs/ — Log files
• *.log, .DS_Store, __pycache__/, *.pyc, node_modules/
• settings.local.json — Local overrides

Important: scripts/completion-state/ (CLI tools) ARE copied. Only the top-level runtime directories are excluded.

.claude/CLAUDE.md Handling

• .claude/CLAUDE.md is harness documentation — ALWAYS overwritten from source
• CLAUDE.md at project root is project-specific — NEVER touched

Runtime Directory Creation

Creates excluded directories with .gitkeep so git tracks the structure:

• state/, progress/, worker-assignments/
• completion-state/ (with subdirs: default/, history/, promises/, sessions/)

.gitignore Updates

Appends Claude Code runtime exclusion entries if not already present.

Git Hook Installation

Installs doc-gardener pre-push hook via attractor/cli.py install-hooks. Skips gracefully for non-git targets or existing non-symlink hooks.

Files Copied vs Excluded

Copied (version controlled)

• settings.json — Core configuration
• skills/ — All skill definitions
• hooks/ — Lifecycle hooks
• output-styles/ — Agent behavior definitions
• scripts/ — CLI utilities (includes scripts/completion-state/)
• commands/ — Slash commands
• schemas/ — JSON schemas
• tests/ — Hook tests
• agents/ — Agent configurations
• documentation/ — Architecture docs
• validation/ — Validation agent configs
• learnings/ — Multi-agent coordination guides

Excluded (runtime, gitignored)

• state/* — Directory kept with .gitkeep
• completion-state/* — Subdirs created: default/, history/, promises/, sessions/
• progress/* — Directory kept with .gitkeep
• worker-assignments/* — Directory kept with .gitkeep
• logs/ — Log files
• settings.local.json — Local overrides

Example Interaction

User: Setup harness in ~/Documents/Windsurf/new-project

Claude: Running deploy-harness.sh --target ~/Documents/Windsurf/new-project

[Script output with verification results]

Harness deployed successfully. 550 files synced. Next steps:
1. Review .claude/CLAUDE.md
2. Commit .claude/ to git
3. Launch: ccsystem3 | ccorch | launchcc

User: Deploy harness to all targets

Claude: Running deploy-harness.sh (deploying to all configured targets)

[Script deploys to my-project and my-project]

Both targets updated successfully.