Claude Flow Integration

Claude-Flow Integration

This project uses a hybrid architecture combining Claude Code's native agent teams with Claude-Flow V3 features. Both systems MUST work together — neither is optional.

Architecture Philosophy

• Priority: Stability and integrity over speed optimization
• Context: Production-quality work, not just internal tooling
• Enforcement: 100% accuracy required for all workflow rules

System Roles

Our Custom System (Blocking Enforcement)

• Agent Teams: TeamCreate + native experimental teams (visible tmux panes, mandatory)
• Hooks: .claude/hooks/*.sh scripts enforce workflow rules (blocking, fail-fast)
• Memory: claude-mem stores human decisions, architectural choices, preferences
• Agent Specs: .claude/agents/*.md define agent capabilities (authoritative for model routing)
• Linear Integration: Single source of truth for requirements, status, tickets

Claude-Flow V3 (Learning & Advisory)

• Hooks: Claude-Flow hooks run AFTER our enforcement hooks (non-blocking, continueOnError: true)
• Background Workers: 10 daemon workers (map, audit, optimize, consolidate, testgaps, ultralearn, deepdive, document, refactor, benchmark)
• Memory: Claude-Flow HNSW memory stores agent patterns, coding learnings, optimization history
• Model Routing: Claude-Flow routing provides suggestions; agent specs are the final authority
• Task Management: Claude-Flow task orchestration supplements native TaskCreate/TaskUpdate
• Production Features: DDD domain tracking, ADR generation, security scanning (CVE, threat modeling)

Serena MCP (Semantic Code Intelligence)

• Symbol Navigation: find_symbol, get_symbols_overview for token-efficient code reading (50-75% savings vs reading full files)
• Impact Analysis: find_referencing_symbols to find all callers/users of any symbol
• Precise Edits: replace_symbol_body, insert_before/after_symbol for symbol-level code modifications
• Rename Refactoring: rename_symbol with automatic reference updates across codebase
• Pattern Search: search_for_pattern for regex search across project files
• Project Memory: read_memory, write_memory for project-specific code structure and conventions

Hook Execution Order (CRITICAL)

Hooks run in this exact order to ensure enforcement before learning:

Write/Edit/MultiEdit:

1. .claude/hooks/protect-hooks.sh (ask — prompts user before hook modifications)
2. .claude/hooks/enforce-orchestrator-delegation-v2.sh (blocking — ensures orchestrator delegates to agents; auto-allows subagents)
3. .claude/hooks/enforce-plan-files.sh (blocking — protects plan/ directory)
4. .claude/hooks/check-unwrap.sh (advisory — warns about .unwrap() in library code)
5. .claude/hooks/scan-secrets.sh (advisory — detects potential secrets/credentials in code)
6. Claude-Flow pre-edit hook (non-blocking — learns patterns, advisory only)

Bash commands:

1. .claude/hooks/protect-hooks.sh (ask — prompts user before hook modifications)
2. .claude/hooks/enforce-orchestrator-delegation-v2.sh (blocking — catches sed/awk/perl bypasses; auto-allows subagents)
3. .claude/hooks/protect-main.sh (blocking — prevents direct main commits)
4. .claude/hooks/enforce-branch-naming.sh (blocking — validates branch format)
5. .claude/hooks/enforce-review-gate.sh (blocking — requires git-review completion)
6. Claude-Flow pre-command hook (non-blocking — learns patterns, advisory only)

Task (agent spawning):

1. .claude/hooks/enforce-visible-agents.sh (blocking — requires team_name for agent visibility)
2. Claude-Flow task hook (non-blocking — advisory only)

Read/Grep/Bash (source files):

1. .claude/hooks/enforce-serena-usage.sh (advisory — suggests Serena for code navigation)

Stop (session end):

1. .claude/hooks/enforce-memory.sh (advisory — reminds to save to claude-mem)
2. .claude/hooks/check-claude-flow-memory.sh (advisory — reminds to save patterns)

TaskCompleted (task marked done):

1. .claude/hooks/enforce-task-quality.sh (blocking — runs cargo test + cargo clippy before task completion; skips non-code tasks)

TeammateIdle (agent going idle):

1. .claude/hooks/enforce-idle-quality.sh (blocking — runs cargo test + cargo clippy before idle; skips if no source changes)

UserPromptSubmit (every prompt):

1. .claude/hooks/enforce-ticket.sh (blocking — requires branch with ticket ID)
2. Claude-Flow routing hook (non-blocking)
3. .claude/hooks/workflow-reminder.sh (advisory)

Memory Separation

• claude-mem: User preferences, architectural decisions, debugging insights, cross-session learnings (human-driven)
• Claude-Flow memory: Agent coordination patterns, code optimization history, model performance data (agent-driven)
• Serena memory: Project code structure, conventions, build commands (code-driven, auto-discovered via onboarding)
• Linear: Ticket requirements, acceptance criteria, status tracking (project management)
• No overlap: Each memory system serves a distinct purpose; never duplicate content

Tool Routing (Serena vs Native Tools)

• Reading code structure: Use Serena get_symbols_overview → only read bodies you need (saves 50-75% tokens)
• Finding definitions: Use Serena find_symbol instead of Grep for symbol definitions
• Finding callers: Use Serena find_referencing_symbols instead of Grep for who-calls-what
• Replacing functions: Use Serena replace_symbol_body for entire function replacement (vs Edit for line-level)
• Non-code files: Use Read/Edit/Grep (Serena only works with LSP-supported languages)
• Config/markdown/TOML: Use Read/Edit (Serena has limited support for non-code)

Model Routing

• Agent specs (.claude/agents/*.md) define the authoritative model for each agent type
• Claude-Flow routing provides suggestions based on task complexity and past performance
• In case of conflict: Agent spec always wins (manual configuration > automated suggestion)
• Model tiers: Opus (3: planner, red-teamer, senior-coder), Sonnet (7: requirements-interviewer, explorer, architect, coder, tech-lead, reviewer, qa), Haiku (4: junior-coder, documentation, explainer, optimizer)
• Example: coder uses Sonnet (per spec), even if Claude-Flow suggests Opus for a task

Background Workers (Always Enabled)

All 10 Claude-Flow daemon workers run continuously:

• map: Codebase structure analysis
• audit: Code quality and security auditing
• optimize: Performance optimization suggestions
• consolidate: Memory and pattern consolidation
• testgaps: Test coverage gap detection
• ultralearn: Advanced pattern learning from agent interactions
• deepdive: Deep analysis of complex code changes
• document: Automatic documentation generation (ADR, DDD docs)
• refactor: Refactoring opportunity detection
• benchmark: Performance benchmarking and regression detection

Production Features (Always Enabled)

• DDD (Domain-Driven Design): Track bounded contexts, validate domain boundaries, maintain /docs/ddd
• ADR (Architecture Decision Records): Auto-generate decision records, maintain /docs/adr, use MADR template
• Security Scanning: Auto-scan on edit, CVE vulnerability checking, threat modeling, security pattern detection

Integration Rules

1. Both systems required: Claude-Flow features supplement native teams; never replace them
2. Hook order matters: Our blocking hooks run first; Claude-Flow hooks learn second
3. Memory separation: Never duplicate data across claude-mem, Claude-Flow memory, and Linear
4. Model authority: Agent specs define models; Claude-Flow routing is advisory only
5. Stability first: All Claude-Flow hooks have continueOnError: true to prevent workflow blockage
6. 100% enforcement: Our custom hooks must maintain perfect accuracy; no false positives

Why This Architecture?

• Complementary systems: Native teams handle coordination; Claude-Flow handles learning
• Fail-safe design: Blocking enforcement prevents mistakes; non-blocking learning improves over time
• Production-ready: DDD, ADR, security features support large-scale professional work
• Clear boundaries: Each system owns specific responsibilities with no overlap
• Observable behavior: tmux panes show agent activity; background workers run invisibly
• Token efficiency: Serena's LSP-powered navigation reads only what's needed (50-75% savings over full-file reads)