Navigator Initialization Skill Skill

Navigator Initialization Skill

Purpose

Creates the Navigator documentation structure (.agent/) in a new project, copies templates, and sets up initial configuration.

When This Skill Auto-Invokes

"Initialize Navigator in this project"
"Set up Navigator documentation structure"
"Create .agent folder for Navigator"
"Bootstrap Navigator for my project"

What This Skill Does

Checks if already initialized: Prevents overwriting existing structure

Creates .agent/ directory structure:

.agent/
├── DEVELOPMENT-README.md
├── .nav-config.json
├── tasks/
├── system/
├── sops/
│   ├── integrations/
│   ├── debugging/
│   ├── development/
│   └── deployment/
└── grafana/
    ├── docker-compose.yml
    ├── prometheus.yml
    ├── grafana-datasource.yml
    ├── grafana-dashboards.yml
    ├── navigator-dashboard.json
    └── README.md

Copies templates: DEVELOPMENT-README.md, config, Grafana setup
Auto-detects project info: Name, tech stack (from package.json if available)
Updates CLAUDE.md: Adds Navigator-specific instructions to project
Creates .gitignore entries: Excludes temporary Navigator files

Execution Steps

1. Check if Already Initialized

if [ -d ".agent" ]; then
    echo "✅ Navigator already initialized in this project"
    echo ""
    echo "To start a session: 'Start my Navigator session'"
    echo "To view documentation: Read .agent/DEVELOPMENT-README.md"
    exit 0
fi

2. Detect Project Information

Read package.json, pyproject.toml, go.mod, Cargo.toml, or similar to extract:

Project name
Tech stack
Dependencies

Fallback: Use current directory name if no config found.

3. Create Directory Structure

Use Write tool to create:

.agent/
.agent/tasks/
.agent/system/
.agent/sops/integrations/
.agent/sops/debugging/
.agent/sops/development/
.agent/sops/deployment/
.agent/grafana/

4. Copy Templates

Copy from plugin's templates/ directory to .agent/:

DEVELOPMENT-README.md:

Replace ${PROJECT_NAME} with detected project name
Replace ${TECH_STACK} with detected stack
Replace ${DATE} with current date

.nav-config.json:

{
  "version": "5.5.0",
  "project_name": "${PROJECT_NAME}",
  "tech_stack": "${TECH_STACK}",
  "project_management": "none",
  "task_prefix": "TASK",
  "team_chat": "none",
  "auto_load_navigator": true,
  "compact_strategy": "conservative",
  "auto_update": {
    "enabled": true,
    "check_interval_hours": 1
  }
}

Grafana Setup: Copy all Grafana dashboard files to enable metrics visualization:

# Find plugin installation directory
PLUGIN_DIR="${HOME}/.claude/plugins/marketplaces/navigator-marketplace"

# Copy Grafana files if plugin has them
if [ -d "${PLUGIN_DIR}/.agent/grafana" ]; then
  cp -r "${PLUGIN_DIR}/.agent/grafana/"* .agent/grafana/
  echo "✓ Grafana dashboard installed"
else
  echo "⚠️  Grafana files not found in plugin"
fi

Files copied:

docker-compose.yml (Grafana + Prometheus stack)
prometheus.yml (scrape config for Claude Code metrics)
grafana-datasource.yml (Prometheus datasource config)
grafana-dashboards.yml (dashboard provider config)
navigator-dashboard.json (10-panel Navigator metrics dashboard)
README.md (setup instructions)

5. Update Project CLAUDE.md

If CLAUDE.md exists:

Append Navigator-specific sections
Keep existing project customizations

If CLAUDE.md doesn't exist:

Copy templates/CLAUDE.md to project root
Customize with project info

6. Claude Code Hooks (Plugin Manifest)

Navigator's lifecycle hooks ship with the plugin manifest (.claude-plugin/plugin.json's top-level hooks field) starting v6.13.0+. They are not merged into the project's .claude/settings.json — Claude Code only substitutes ${CLAUDE_PLUGIN_DIR} for hooks declared in a plugin manifest, so merging them into user settings (the prior approach) produced broken commands like /hooks/X.py.

The plugin registers the following hooks automatically when the plugin is installed:

SessionStart — injects Navigator context (navigator + active marker + config + graph + profile) into the session, eliminating ~6 Read tool calls at start (v6.9.0+)
PreCompact — writes a context marker before manual or silent auto-compact (v6.10.0+)
PostCompact — appends Claude Code's compact summary to the marker (v6.10.0+)
Stop — silent workflow-state writer; records whether WORKFLOW CHECK / NAVIGATOR_STATUS appeared (v6.11.0+)
UserPromptSubmit — workflow_enforcer (Loop/Task mode trigger detection + optional strict_block gate)
PreToolUse(Read) — nav_read_guard (bulk-read circuit breaker)
PostToolUse(Edit|Write|Bash) — token monitor (warns at 70% / 85% context usage)
PostToolUse(Edit|Write) — task→graph sync and profile correction sync (v6.11.0+)

Nothing for nav-init to do here. The skill no longer writes to .claude/settings.json for hooks.

⚠️  RESTART REQUIRED to activate hooks after plugin install/update.
   Claude Code caches plugin manifest hooks at session start.

Opt-out: Users can disable any hook via .agent/.nav-config.json:

{
  "session_start_hook":    { "enabled": false },
  "compact_hook":          { "enabled": false },
  "workflow_state_hook":   { "enabled": false },
  "task_graph_sync_hook":  { "enabled": false },
  "profile_sync_hook":     { "enabled": false },
  "workflow_enforcer_hook":{ "enabled": false }
}

(The hooks themselves read this config and short-circuit when disabled, so no settings.json mutation is needed.)

7. Create .gitignore Entries

Add to .gitignore if not present:

# Navigator context markers
.context-markers/

# Navigator temporary files
.agent/.nav-temp/

8. Success Message

✅ Navigator Initialized Successfully!

Created structure:
  📁 .agent/                    Navigator documentation
  📁 .agent/tasks/              Implementation plans
  📁 .agent/system/             Architecture docs
  📁 .agent/sops/               Standard procedures
  📁 .agent/grafana/            Metrics dashboard
  📄 .agent/.nav-config.json    Configuration
  📄 CLAUDE.md                  Updated with Navigator workflow

Next steps:
  1. Start session: "Start my Navigator session"
  2. Optional: Enable metrics - see .agent/sops/integrations/opentelemetry-setup.md
  3. Optional: Launch Grafana - cd .agent/grafana && docker compose up -d

Token monitoring is active - you'll be warned when approaching context limits.

Documentation: Read .agent/DEVELOPMENT-README.md

Error Handling

If .agent/ exists:

Don't overwrite
Show message: "Already initialized"

If templates not found:

Error: "Navigator plugin templates missing. Reinstall plugin."

If no write permissions:

Error: "Cannot create .agent/ directory. Check permissions."

Predefined Functions

`project_detector.py`

def detect_project_info(cwd: str) -> dict:
    """
    Detect project name and tech stack from config files.

    Checks (in order):
    1. package.json (Node.js)
    2. pyproject.toml (Python)
    3. go.mod (Go)
    4. Cargo.toml (Rust)
    5. composer.json (PHP)
    6. Gemfile (Ruby)

    Returns:
        {
            "name": "project-name",
            "tech_stack": "Next.js, TypeScript, Prisma",
            "detected_from": "package.json"
        }
    """

`template_customizer.py`

def customize_template(template_content: str, project_info: dict) -> str:
    """
    Replace placeholders in template with project-specific values.

    Placeholders:
    - ${PROJECT_NAME}
    - ${TECH_STACK}
    - ${DATE}
    - ${YEAR}

    Returns customized template content.
    """

`settings_merger.py`

def merge(target_path: Path, fragment: dict) -> dict:
    """
    Idempotent JSON merge for .claude/settings.json.

    - If target doesn't exist: create from fragment.
    - If target exists: deep-merge `hooks` arrays by event name; dedupe entries
      by command string. Preserves user-defined hooks and other top-level keys.
    - Refuses to clobber invalid JSON (exits 2).
    """

Status (v6.13.0+): Retained as general-purpose JSON merger for non-hook keys (permissions, mcpServers, etc.) that downstream skills may want to preserve. Navigator no longer passes a hooks fragment through it — hooks now ship with the plugin manifest (.claude-plugin/plugin.json). The hook-merging code path is dead-for-Navigator but kept intact for safety (existing call sites that already pass non-hook fragments continue to work).

Examples

Example 1: New Next.js Project

User says: "Initialize Navigator in this project"

Skill detects:

package.json exists
Name: "my-saas-app"
Dependencies: next, typescript, prisma

Result:

.agent/ created
DEVELOPMENT-README.md shows: "Project: My SaaS App"
DEVELOPMENT-README.md shows: "Tech Stack: Next.js, TypeScript, Prisma"
.nav-config.json has project_name: "my-saas-app"

Example 2: Python Project

User says: "Set up Navigator"

Skill detects:

pyproject.toml exists
Name: "ml-pipeline"
Dependencies: fastapi, pydantic, sqlalchemy

Result:

.agent/ created
Tech stack: "FastAPI, Pydantic, SQLAlchemy"

Example 3: Already Initialized

User says: "Initialize Navigator"

Skill checks:

.agent/ directory exists

Result:

✅ Navigator already initialized in this project

To start a session: 'Start my Navigator session'

Integration with Other Skills

nav-start skill:

Checks for .agent/DEVELOPMENT-README.md
If missing, suggests: "Initialize Navigator first"

nav-task skill:

Creates tasks in .agent/tasks/
Requires initialization

nav-sop skill:

Creates SOPs in .agent/sops/
Requires initialization

Version History

1.0.0 (2025-01-20): Initial implementation
- Auto-detection of project info
- Template customization
- Grafana setup included
- Error handling for existing installations

Notes

This skill replaces the deleted /nav:init command from v2.x
Templates are copied from plugin installation directory
Project info detection is best-effort (falls back to directory name)
Safe to run multiple times (won't overwrite existing structure)

Agent Skills: Navigator Initialization Skill

Install this agent skill to your local

Skill Files