bkit-rules（Bkit 规则）

bkit Core Rules

Automatically applied rules that don't require user commands.

1. PDCA Auto-Apply Rules

No Guessing: If unsure, check docs → If not in docs, ask user SoR Priority: Code > CLAUDE.md > docs/ design documents

Template References

2. Level Auto-Detection

Detection Order

1. Check CLAUDE.md for explicit Level declaration
2. File structure based detection

Enterprise (2+ conditions met)

• infra/terraform/ folder
• infra/k8s/ or kubernetes/ folder
• services/ folder (2+ services)
• turbo.json or pnpm-workspace.yaml
• docker-compose.yml
• .github/workflows/ (CI/CD)

Dynamic (1+ conditions met)

• bkend settings in.mcp.json
• lib/bkend/ or src/lib/bkend/
• supabase/ folder
• firebase.json

Starter

None of the above conditions met.

Level-specific Behavior

Level Upgrade Signals

• Starter → Dynamic: "Add login", "Save data", "Admin page"
• Dynamic → Enterprise: "High traffic", "Microservices", "Own server"

Hierarchical CLAUDE.md Rules

project/
├── CLAUDE.md                 # Project-wide (always reference)
├── services/CLAUDE.md        # Backend work context
├── frontend/CLAUDE.md        # Frontend work context
└── infra/CLAUDE.md           # Infrastructure context

Rule: Area-specific rules > Project-wide rules

3. Agent Auto-Trigger Rules

Level-Based Selection

When user requests feature development:

1. Detect project level
2. Invoke appropriate agent automatically

Task-Based Selection

Proactive Suggestions

After completing major tasks, suggest relevant agents.

Do NOT Auto-Invoke When

• User explicitly declines
• Task is trivial
• User wants to understand process
• Agent already invoked for same task

4. Code Quality Standards

Pre-coding Checks

1. Does similar functionality exist? Search first
2. Check utils/, hooks/, components/ui/
3. Reuse if exists; create if not

Core Principles

DRY: Extract to common function on 2nd use SRP: One function, one responsibility No Hardcoding: Use meaningful constants Extensibility: Write in generalized patterns

Self-Check After Coding

• Same logic exists elsewhere?
• Can function be reused?
• Hardcoded values present?
• Function does only one thing?

When to Refactor

• Same code appears 2nd time
• Function exceeds 20 lines
• if-else nests 3+ levels
• Same parameters passed to multiple functions

5. Task Classification

Classify tasks to apply appropriate PDCA level:

Classification Keywords

Quick Fix: fix, typo, correct, adjust, tweak Minor Change: improve, refactor, enhance, optimize, update Feature: add, create, implement, build, new feature Major Feature: redesign, migrate, architecture, overhaul, rewrite

6. Output Style Auto-Selection (v1.5.1)

When project level is detected, automatically suggest the matching output style:

Auto-Selection Rules

• On session start: Suggest output style matching detected level
• On /starter init, /dynamic init, /enterprise init: Auto-suggest style for that level
• On PDCA phase transitions: Suggest bkit-pdca-guide if not already active
• User can override with /output-style at any time

Available Output Styles

7. Agent Teams Auto-Suggestion (v1.5.1)

Suggest Agent Teams when conditions are met:

Suggestion Triggers

Team Availability

Requirements

• Environment: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
• If env var not set: Suggest setting it when team mode would be beneficial
• Command: /pdca team {feature} to start team mode

8. Agent Memory Awareness (v1.5.1)

Agent Memory is automatically active for all bkit agents. No user action required.

How It Works

• Agents remember project context across sessions via memory: project scope
• Some agents (starter-guide, pipeline-guide) use memory: user for cross-project learning
• Memory persists in .claude/agent-memory/ (project) or ~/.claude/agent-memory/ (user)

Memory Scopes

Proactive Mention

• On session start: "Agent Memory is active — agents remember context across sessions"
• When agent is invoked: Agent may reference previous session context
• No configuration needed — fully automatic

9. Plugin Hot Reload (v1.6.0)

After modifying bkit plugin files, use /reload-plugins to apply changes without restarting Claude Code.

• No need to exit and re-enter the session
• Changes to skills, agents, hooks, and templates are reflected immediately
• Note: Changes to CLAUDE.md require /clear to fully refresh

Wildcard Permissions (CC 2.1.0+)

CC 2.1.0+ supports Bash(pattern*) wildcard permissions.

bkit Recommended Patterns

• Bash(npm *) - Allow all npm commands
• Bash(git log*) - Allow git log variants
• Bash(node *) - Allow node execution
• Bash(npx *) - Allow npx execution

Deny Recommendations

• Bash(rm -rf*) - Deny recursive delete (dangerous)
• Bash(git push --force*) - Deny force push

Configuration

Add to .claude/settings.json:

{
  "permissions": {
    "allow": ["Bash(npm *)", "Bash(git log*)"],
    "deny": ["Bash(rm -rf*)"]
  }
}