flock-core 0.5.11__tar.gz → 0.5.21__tar.gz

flock_core-0.5.21/AGENTS.md

@@ -0,0 +1,1789 @@
+# AGENTS.md
+
+**Welcome, AI coding agent!** 👋
+
+This is Flock, a production-grade blackboard-first AI agent orchestration framework. This guide gets you up to speed quickly on the current project state and development patterns.
+
+**Current Version:** Backend: 0.5.0b63 • Frontend: 0.1.4
+**Architecture:** Hybrid Python/TypeScript with real-time dashboard
+**Package Manager:** UV (NOT pip!)
+**Status:** Production-ready with comprehensive monitoring
+
+---
+
+## 🎯 Project Snapshot
+
+### What Is Flock?
+
+A blackboard architecture framework where specialized AI agents collaborate through a shared typed workspace—no direct coupling, no rigid workflows, just emergent intelligence.
+
+**Core Pattern:** Blackboard Architecture (like Hearsay-II from 1970s, but for modern LLMs)
+
+**Key Differentiator:** The only framework treating blackboard orchestration as a first-class citizen with built-in visibility controls, real-time monitoring, and enterprise-grade safety features.
+
+### Architecture in 30 Seconds
+
+```
+┌─────────────────────────────────────────────┐
+│         Blackboard (Typed Artifacts)        │
+│  ┌──────┐ → ┌──────┐ → ┌──────┐ → ┌──────┐│
+│  │ Idea │   │Movie │   │Script│   │Review││
+│  └──────┘   └──────┘   └──────┘   └──────┘│
+└─────────────────────────────────────────────┘
+      ↑           ↑           ↑           ↑
+  Agent A     Agent B     Agent C     Agent D
+  (produce)   (consume    (consume    (consume
+              & produce)  & produce)  & produce)
+```
+
+**Key Concepts:**
+- **Artifacts:** Typed data (Pydantic models) published to blackboard
+- **Subscriptions:** Declarative rules for when agents react
+- **Visibility:** Built-in access control (Public/Private/Tenant/Label-based/Time-based)
+- **Fan-Out Publishing:** Produce multiple artifacts from single agent execution with filtering/validation
+- **Components:** Three levels of extensibility:
+  - **Orchestrator Components:** Global lifecycle hooks (monitoring, metrics, coordination)
+  - **Agent Components:** Per-agent behavior (quality gates, retry logic, validation)
+  - **Engines:** Custom processing logic (DSPy, regex, deterministic rules)
+- **Real-time Dashboard:** React/TypeScript interface for live monitoring
+
+---
+
+## 🚀 Quick Setup
+
+### Prerequisites
+
+- **Python 3.12+** (we use modern async features)
+- **UV package manager** (faster than pip, handles virtual envs)
+- **Node.js 18+** (22+ recommended) for dashboard frontend
+- **OpenAI API key** (for running examples)
+
+### Installation
+
+```bash
+# Clone repo
+git clone https://github.com/yourusername/flock-flow.git
+cd flock-flow
+
+# Install Python dependencies (UV creates venv automatically)
+poe install  # Equivalent to: uv sync --dev --all-groups --all-extras
+
+# Set up environment
+export OPENAI_API_KEY="sk-..."
+export DEFAULT_MODEL="openai/gpt-4.1"
+
+# Verify installation
+uv run python -c "from flock import Flock; print('✅ Ready!')"
+```
+
+### Run Examples
+
+```bash
+# CLI examples (with detailed output)
+uv run python examples/01-cli/01_declarative_pizza.py
+uv run python examples/01-cli/02_input_and_output.py
+uv run python examples/01-cli/03_code_detective.py
+
+# Dashboard examples (with visualization)
+uv run python examples/02-dashboard/01_declarative_pizza.py
+uv run python examples/02-dashboard/02_input_and_output.py
+uv run python examples/02-dashboard/03_code_detective.py
+
+# Engine + component playgrounds
+uv run python examples/05-engines/emoji_mood_engine.py
+uv run python examples/05-engines/potion_batch_engine.py
+uv run python examples/06-agent-components/plot_twist_component.py
+uv run python examples/06-agent-components/cheer_meter_component.py
+uv run python examples/07-orchestrator-components/quest_tracker_component.py
+uv run python examples/07-orchestrator-components/kitchen_monitor_component.py
+```
+
+---
+
+## 📚 Detailed Guides
+
+For deep dives into specific topics, see:
+
+**Core Architecture:**
+- **[Architecture Overview](docs/architecture.md)** - Complete system design, module organization, and Phase 1-7 refactoring ⭐ **UPDATED**
+- **[Architecture & Blackboard](docs/guides/blackboard.md)** - Core pattern, structure, and behavior
+- **[Agent Guide](docs/guides/agents.md)** - Complete agent development reference
+- **[Context Providers](docs/guides/context-providers.md)** - Smart filtering & security boundaries for agent context ⭐ **NEW in 0.5**
+
+**Code Patterns & Standards:**
+- **[Error Handling Patterns](docs/patterns/error_handling.md)** - Production-ready error handling guide ⭐ **NEW**
+- **[Async Patterns](docs/patterns/async_patterns.md)** - Async/await best practices and common pitfalls ⭐ **NEW**
+
+**Components & Extensibility:**
+- **[Agent Components](docs/guides/components.md)** - Extend agent behavior with lifecycle hooks
+- **[Orchestrator Components](docs/guides/orchestrator-components.md)** - Extend orchestrator behavior with lifecycle hooks (NEW!)
+
+**Logic Operations (Advanced Subscriptions):**
+- **[Predicates](docs/guides/predicates.md)** - Conditional consumption with `where=` filters
+- **[Join Operations](docs/guides/join-operations.md)** - Correlate related artifacts with JoinSpec
+- **[Batch Processing](docs/guides/batch-processing.md)** - Efficient bulk operations with BatchSpec
+
+**Publishing Patterns:**
+- **[Fan-Out Publishing](docs/guides/fan-out.md)** - Generate multiple outputs with filtering/validation ⭐ **NEW in 0.5**
+
+**Development & Operations:**
+- **[Development Workflow](docs/guides/testing.md)** - Testing, quality, versioning, pre-commit
+- **[Frontend/Dashboard](docs/guides/dashboard.md)** - Dashboard usage and development
+- **[Configuration & Dependencies](CONTRIBUTING.md)** - Environment and setup
+- **[Patterns & Common Tasks](docs/guides/patterns.md)** - Recipes, performance, security
+- **[Development Workflow](docs/about/contributing.md)** - Testing, quality, versioning, pre-commit
+- **[Frontend/Dashboard](docs/guides/dashboard.md)** - Dashboard usage and development
+- **[Configuration & Dependencies](docs/reference/configuration.md)** - Environment and setup
+- **[Patterns & Common Tasks](docs/guides/patterns.md)** - Recipes, performance, security
+
+---
+
+## 🚨 CRITICAL PATTERNS (Learn from Our Experience)
+
+### ⚡ invoke() vs run_until_idle() - Execution Control (no double-run by default)
+
+**This pattern cost us hours of debugging - learn from our pain!**
+
+#### The Goal
+Choose between isolated agent execution and event-driven cascades.
+
+#### How it works
+`invoke()` behavior depends on `publish_outputs`:
+
+```python
+# Isolated: execute agent only (no cascade)
+await orchestrator.invoke(agent, input_artifact, publish_outputs=False)
+
+# Cascading: execute and publish outputs, then run downstream agents
+await orchestrator.invoke(agent, input_artifact, publish_outputs=True)
+await orchestrator.run_until_idle()  # processes downstream agents
+```
+
+Notes:
+- By default, agents have `prevent_self_trigger=True`, so an agent will not re-run on its own outputs when you call `run_until_idle()`. Downstream agents that subscribe to the output types will run.
+
+#### When to Use Which Pattern
+
+**✅ USE `publish_outputs=False` for:**
+- Unit testing specific agent behavior
+- Testing component hooks in isolation
+- Direct execution without cascade
+- Most test scenarios
+
+```python
+# ✅ CORRECT: Test component execution order
+await orchestrator.invoke(agent, input_artifact, publish_outputs=False)
+assert component_order == ["A", "B", "C"]
+```
+
+**✅ USE `publish_outputs=True` + `run_until_idle()` for:**
+- Integration testing agent cascades
+- Testing multi-agent workflows
+- End-to-end scenario validation
+- Event-driven behavior testing
+
+```python
+# ✅ CORRECT: Test agent cascade
+await orchestrator.invoke(agent_a, input_artifact, publish_outputs=True)
+await orchestrator.run_until_idle()  # Process agent_b, agent_c...
+assert len(output_artifacts) == 3
+```
+
+#### Quick Reference
+
+| Scenario | invoke() call | run_until_idle() | Result |
+|----------|---------------|------------------|--------|
+| Unit test | `invoke(..., publish_outputs=False)` | No | Single execution |
+| Integration test | `invoke(..., publish_outputs=True)` | Yes | Cascade to downstream agents |
+| Common mistake | `invoke(..., publish_outputs=True)` | Yes | Not a double-run; downstream agents run |
+
+Rule of thumb: start with `publish_outputs=False` for unit tests; enable publication only when you want cascades.
+
+---
+
+### ⚡ Batching Pattern: publish() + run_until_idle() Separation = Parallel Power
+
+**Why `run_until_idle()` is separate from `publish()` - it's not a bug, it's a feature!**
+
+#### Sequential vs Parallel Execution
+
+**❌ If `run_until_idle()` was built into `publish()`:**
+```python
+# Hypothetical auto-run design
+await flock.publish(review1)  # Publishes AND waits for completion
+await flock.publish(review2)  # Publishes AND waits for completion
+await flock.publish(review3)  # Publishes AND waits for completion
+
+# Result: SEQUENTIAL processing (3x time)
+```
+
+**✅ With current design (separated):**
+```python
+# Queue up multiple artifacts
+await flock.publish(review1)  # Schedules agents
+await flock.publish(review2)  # Schedules agents
+await flock.publish(review3)  # Schedules agents
+
+# Now trigger execution
+await flock.run_until_idle()  # All independent agents run in PARALLEL!
+
+# Result: PARALLEL processing (~1x time if agents are independent)
+```
+
+#### Best Practices
+
+**✅ DO: Batch when possible**
+```python
+# Good: Batch-publish customer reviews
+for review in customer_reviews:
+    await flock.publish(review)
+await flock.run_until_idle()
+```
+
+**✅ DO: Use for multi-type workflows**
+```python
+# Good: Publish different types, let agents run in parallel
+await flock.publish(XRayImage(...))
+await flock.publish(LabResults(...))
+await flock.publish(PatientHistory(...))
+await flock.run_until_idle()  # Radiologist, lab_tech, historian run concurrently
+```
+
+**⚠️ CAREFUL: Separate workflows with traced_run()**
+```python
+# Better: Separate workflows explicitly
+async with flock.traced_run("review_workflow"):
+    await flock.publish(review_workflow_input)
+    await flock.run_until_idle()
+
+async with flock.traced_run("order_workflow"):
+    await flock.publish(order_workflow_input)
+    await flock.run_until_idle()
+```
+
+#### Key Takeaway
+
+The separation of `publish()` and `run_until_idle()` gives you **control over execution timing and batching**. This enables:
+- ⚡ **Parallel execution** when agents are independent
+- 🎛️ **Fine-grained control** over when execution happens
+- 📊 **Better performance** for bulk operations
+- 🔍 **Clearer workflow boundaries** with `traced_run()`
+
+**This is not a bug or oversight - it's a fundamental design choice that enables patterns other frameworks can't easily support.**
+
+---
+
+### 🗄️ Persistent Blackboard History (SQLite Store)
+
+**Production teams asked for auditability—now the blackboard can keep a full historical trail.**
+
+#### Why this matters
+- **Durable artifacts**: Everything on the blackboard (payloads, tags, visibility, correlation IDs) is stored on disk for replay, compliance, and debugging.
+- **Faster postmortems**: Query `/api/v1/artifacts` with filters (`type`, `produced_by`, `tags`, `visibility`, time windows) to reconstruct agent cascades after the fact.
+- **Operational dashboards**: The new **Historical Blackboard** module preloads persisted pages, exposes consumption metadata, and shows retention banners so operators know how far back they can scroll.
+- **Retention policies**: CLI helpers make maintenance routine—`flock sqlite-maintenance my.db --delete-before 2025-01-01T00:00:00Z --vacuum` prunes old artifacts and compacts the store.
+
+#### Quick start
+```python
+from flock import Flock
+from flock.store import SQLiteBlackboardStore
+
+store = SQLiteBlackboardStore(".flock/history.db")
+await store.ensure_schema()
+
+flock = Flock("openai/gpt-4.1", store=store)
+await flock.publish(MyDreamPizza(pizza_idea="fermented garlic delight"))
+await flock.run_until_idle()
+```
+
+Kick the tyres with `examples/02-the-blackboard/01_persistent_pizza.py`, then launch `examples/03-the-dashboard/04_persistent_pizza_dashboard.py` to inspect the retained history alongside live WebSocket updates.
+
+> **Heads-up:** The interface now returns `ArtifactEnvelope` objects when `embed_meta=True`. Future backends (Postgres, BigQuery, etc.) can implement the same contract to plug straight into the runtime and dashboard.
+
+---
+
+### 🔐 Context Providers - Smart Filtering & Security Boundaries
+
+**Context Providers are the intelligent filter layer between agents and the blackboard—controlling what each agent sees, reducing token costs by 90%+, and automatically redacting sensitive data.**
+
+**⚠️ IMPORTANT FOR AI AGENTS:** Context Providers implement the security architecture described in [docs/architecture.md](docs/architecture.md) and follow error handling patterns from [docs/patterns/error_handling.md](docs/patterns/error_handling.md). Read these documents to understand the design principles.
+
+#### Why Context Providers Matter
+
+**The Problem:**
+- Agents seeing ALL artifacts wastes tokens (costs)
+- No way to filter sensitive data before agents see it
+- Performance degrades as blackboard grows
+- Security vulnerabilities if agents bypass visibility
+
+**The Solution:**
+Context Providers enforce a security boundary that:
+- ✅ **Reduces token costs** - Agents see only relevant artifacts (90%+ savings)
+- ✅ **Protects sensitive data** - Auto-redact passwords, API keys, credit cards
+- ✅ **Improves performance** - Less context = faster agent execution
+- ✅ **Enforces security** - Agents cannot bypass provider filtering (architecturally impossible!)
+
+**Understanding the Three-Layer Model:**
+
+Flock uses three complementary filtering layers:
+
+1. **Visibility** (security boundary) - Controls BOTH which agents trigger AND which artifacts they see in context
+2. **Subscription Filters** (routing logic) - Controls WHEN agents trigger (e.g., `.consumes(Task, tags={"urgent"})`)
+3. **Context Providers** (context shaping) - Controls WHAT agents see in their historical context
+
+These layers work together to provide fine-grained control over agent execution and data access. Context Providers are the third layer—they filter what agents see in context but do NOT control triggering.
+
+#### Quick Start: Global Filtering
+
+**Filter all agents to see only urgent items:**
+
+```python
+from flock import Flock
+from flock.context_provider import FilteredContextProvider
+from flock.store import FilterConfig
+
+# Create provider that filters by tags
+urgent_only = FilteredContextProvider(
+    FilterConfig(tags={"urgent", "high"}),
+    limit=50  # Max 50 artifacts
+)
+
+# Apply to ALL agents
+flock = Flock("openai/gpt-4.1", context_provider=urgent_only)
+
+# Publish with tags
+await flock.publish(Task(...), tags={"urgent"})  # ✅ Agents see this
+await flock.publish(Task(...), tags={"low"})     # ❌ Agents don't see this
+```
+
+#### Per-Agent Filtering (Role-Based)
+
+**Different agents see different context:**
+
+```python
+# Global: Errors only
+error_provider = FilteredContextProvider(FilterConfig(tags={"ERROR"}))
+flock = Flock("openai/gpt-4.1", context_provider=error_provider)
+
+# Junior engineer: Errors only (uses global)
+junior = flock.agent("junior").consumes(LogEntry).publishes(Report).agent
+
+# Senior engineer: Errors + Warnings (override global)
+warn_provider = FilteredContextProvider(FilterConfig(tags={"ERROR", "WARN"}))
+senior = flock.agent("senior").consumes(LogEntry).publishes(Analysis).agent
+senior.context_provider = warn_provider  # 🎯 Per-agent override!
+
+# Platform team: Everything (override global)
+all_provider = FilteredContextProvider(FilterConfig(tags={"DEBUG", "INFO", "WARN", "ERROR"}))
+platform = flock.agent("platform").consumes(LogEntry).publishes(Report).agent
+platform.context_provider = all_provider
+```
+
+**Provider Priority:**
+```
+Per-Agent Provider  >  Global Provider  >  DefaultContextProvider
+     (highest)            (medium)              (fallback)
+```
+
+#### Production-Ready: Password Redaction
+
+**Automatically redact sensitive data from agent context:**
+
+```python
+from examples.context_provider import PasswordRedactorProvider
+
+# Create provider with built-in patterns
+provider = PasswordRedactorProvider(
+    redaction_text="[REDACTED]",
+    redact_emails=True,
+    log_redactions=True  # Audit trail
+)
+
+flock = Flock("openai/gpt-4.1", context_provider=provider)
+
+# Publish data with secrets
+await flock.publish(UserData(
+    password="MySecret123",           # → [REDACTED]
+    api_key="sk-1234567890abcdef",    # → [REDACTED]
+    email="user@example.com"          # → [REDACTED]
+))
+
+# Agents see redacted version automatically!
+```
+
+**What gets redacted:**
+- 🔒 Passwords and secrets
+- 🔑 API keys (OpenAI, AWS, GitHub, GitLab, Stripe, Google)
+- 🎫 Bearer tokens and JWT
+- 💳 Credit card numbers (Visa, MC, Amex, Discover)
+- 🆔 Social Security Numbers
+- 🔐 Private keys (RSA, EC, OpenSSH)
+- 📧 Email addresses (optional)
+
+#### FilterConfig API Reference
+
+**Declarative filtering with multiple criteria:**
+
+```python
+from flock.store import FilterConfig
+
+# Combine multiple filters (AND logic)
+config = FilterConfig(
+    tags={"ERROR", "CRITICAL"},           # OR logic: any of these tags
+    type_names={"SystemEvent", "Alert"},  # OR logic: any of these types
+    produced_by=["monitor", "watchdog"],  # OR logic: from any of these agents
+    correlation_id="incident-123"         # Exact match
+)
+
+provider = FilteredContextProvider(config, limit=100)
+```
+
+**Criteria are combined with AND logic:**
+- Must have (ERROR OR CRITICAL) tag
+- AND must be (SystemEvent OR Alert) type
+- AND must be from (monitor OR watchdog) agent
+- AND must have correlation_id "incident-123"
+
+#### Custom Context Providers
+
+**Build your own filtering logic:**
+
+```python
+from flock.context_provider import BaseContextProvider, ContextRequest
+from flock.artifacts import Artifact
+
+class TimeBoundProvider(BaseContextProvider):
+    """Only show artifacts from last 1 hour.
+
+    Security is enforced automatically by BaseContextProvider!
+    """
+
+    async def get_artifacts(self, request: ContextRequest) -> list[Artifact]:
+        cutoff = datetime.now() - timedelta(hours=1)
+
+        # Query all artifacts
+        artifacts, _ = await request.store.query_artifacts(limit=1000)
+
+        # Apply time-based filtering
+        recent = [a for a in artifacts if a.created_at >= cutoff]
+
+        return recent
+
+        # ✨ BaseContextProvider automatically:
+        # - Filters by visibility (MANDATORY, cannot bypass!)
+        # - Excludes requested IDs
+        # - Serializes to standard format
+
+# Usage
+provider = TimeBoundProvider()
+flock = Flock("openai/gpt-4.1", context_provider=provider)
+```
+
+**Why this is better:**
+- ✅ **75% less code** - 5 lines vs 20+ lines
+- ✅ **Impossible to bypass security** - visibility enforced by base class
+- ✅ **Consistent serialization** - automatic standard format
+- ✅ **Focus on business logic** - not infrastructure
+
+#### Real-World Use Cases
+
+**Healthcare (HIPAA Compliance):**
+```python
+# Redact PHI (Protected Health Information)
+from examples.context_provider import PasswordRedactorProvider
+
+phi_provider = PasswordRedactorProvider(
+    custom_patterns={
+        "ssn": r"\b\d{3}-\d{2}-\d{4}\b",
+        "mrn": r"MRN[:\s]*\d{6,10}",
+        "dob": r"\b\d{2}/\d{2}/\d{4}\b",
+    }
+)
+
+flock = Flock("openai/gpt-4.1", context_provider=phi_provider)
+```
+
+**DevOps (Log Filtering):**
+```python
+# Error engineers see only errors
+error_provider = FilteredContextProvider(FilterConfig(tags={"ERROR", "CRITICAL"}))
+error_agent.context_provider = error_provider
+
+# Platform team sees everything
+all_provider = FilteredContextProvider(FilterConfig(tags={"DEBUG", "INFO", "WARN", "ERROR"}))
+platform_agent.context_provider = all_provider
+```
+
+**Multi-Tenant SaaS (Customer Isolation):**
+```python
+# Each tenant gets isolated context
+tenant_a_provider = FilteredContextProvider(FilterConfig(correlation_id=f"tenant_{customer_a.id}"))
+tenant_a_agent.context_provider = tenant_a_provider
+
+tenant_b_provider = FilteredContextProvider(FilterConfig(correlation_id=f"tenant_{customer_b.id}"))
+tenant_b_agent.context_provider = tenant_b_provider
+```
+
+#### Performance Impact
+
+**Token Cost Reduction:**
+
+```python
+# ❌ Before (no filtering): Agent sees 1000 artifacts = 200,000 tokens
+# Cost: $0.006 per agent call
+
+# ✅ After (smart filtering): Agent sees 10 relevant artifacts = 2,000 tokens
+# Cost: $0.00006 per agent call
+
+# 💰 99% cost reduction!
+
+provider = FilteredContextProvider(FilterConfig(tags={"urgent"}), limit=10)
+```
+
+#### Security Best Practices
+
+**✅ DO:**
+- Always enforce visibility (NEVER skip)
+- Use `FilteredContextProvider` for common patterns
+- Test redaction patterns with real data
+- Log security events for audit trails
+- Limit by default (start restrictive)
+
+**❌ DON'T:**
+- Don't skip visibility filtering (security boundary!)
+- Don't leak sensitive data in nested objects
+- Don't cache redacted data without re-applying redaction
+- Don't trust agent input (agents are untrusted)
+- Don't expose infrastructure details
+
+#### Learn More
+
+**📚 Complete Guide:** [docs/guides/context-providers.md](docs/guides/context-providers.md) - Architecture, security model, advanced patterns
+
+**💡 Examples:** [examples/08-context-provider/](examples/08-context-provider/) - 5 progressive examples from beginner to expert
+
+**🎁 Production Code:** [examples/08-context-provider/05_password_redactor.py](examples/08-context-provider/05_password_redactor.py) - Copy-paste ready password filtering
+
+#### Quick Reference
+
+```python
+# Global filtering
+flock = Flock("openai/gpt-4.1", context_provider=FilteredContextProvider(FilterConfig(tags={"urgent"})))
+
+# Per-agent override
+agent.context_provider = FilteredContextProvider(FilterConfig(tags={"ERROR", "WARN"}))
+
+# Password redaction
+from examples.context_provider import PasswordRedactorProvider
+flock = Flock("openai/gpt-4.1", context_provider=PasswordRedactorProvider())
+
+# Custom provider
+class MyProvider(ContextProvider):
+    async def __call__(self, request: ContextRequest) -> list[dict[str, Any]]:
+        # Query + filter by visibility (MANDATORY) + your logic
+        pass
+```
+
+---
+
+### 🔒 Test Isolation and Mock Cleanup - The Contamination Trap
+
+**We fixed 32 failing tests caused by test contamination - here's what we learned:**
+
+#### The Problem
+Tests were modifying class-level properties with PropertyMock that persisted across test boundaries:
+
+```python
+# ❌ WRONG: Contaminates other tests
+def test_something(orchestrator):
+    type(orchestrator).agents = PropertyMock(return_value=[mock_agent])
+    # No cleanup - mock persists!
+```
+
+#### The Solution
+Always use fixture cleanup with yield pattern:
+
+```python
+# ✅ CORRECT: Proper cleanup
+@pytest.fixture
+def dashboard_service_with_mocks(orchestrator):
+    original_agents = getattr(type(orchestrator), "agents", None)
+    type(orchestrator).agents = PropertyMock(return_value=[mock_agent])
+    try:
+        yield service
+    finally:
+        # Restore original or delete if it didn't exist
+        if original_agents is not None:
+            type(orchestrator).agents = original_agents
+        elif hasattr(type(orchestrator), "agents"):
+            delattr(type(orchestrator), "agents")
+```
+
+#### Best Practices for Test Isolation
+
+1. **Create Helper Functions for Complex Mocks**
+2. **Use Fixture Cleanup Pattern** (store original, try/finally, restore)
+3. **Test in Isolation First** (run files individually to check contamination)
+4. **Common Contamination Sources**: PropertyMock on class attributes, module-level patches, shared mutable state, async event loops, Rich/logging state pollution
+
+---
+
+### 📦 Version Bumping - Don't Break PyPI Releases!
+
+**⚠️ CRITICAL FOR ALL CODE CHANGES:** Always increment version numbers when making changes that will be committed and pushed. Forgetting this breaks the PyPI publishing workflow!
+
+#### Why This Matters
+
+The automated PyPI release pipeline checks if the version in `pyproject.toml` has been incremented. If you push code changes without bumping the version:
+- ❌ PyPI publish workflow fails
+- ❌ Users can't get your fixes/features via `pip install`
+- ❌ Other developers get confused about which version has which features
+
+#### What to Bump
+
+**Backend version (REQUIRED for backend code changes):**
+```toml
+# pyproject.toml
+[project]
+version = "0.5.0b56"  # Increment this! e.g., "0.5.0b57"
+```
+
+**Frontend version (REQUIRED for dashboard/UI changes):**
+```json
+// src/flock/frontend/package.json
+{
+  "version": "0.1.2"  // Increment this! e.g., "0.1.3"
+}
+```
+
+#### When to Bump
+
+**✅ ALWAYS bump backend version for:**
+- Any Python code changes in `src/flock/`
+- Bug fixes, features, refactors
+- Commits that modify backend behavior
+
+Docs-only changes (in `docs/`, `README.md`) do not require a version bump.
+
+**✅ ALWAYS bump frontend version for:**
+- Dashboard UI changes
+- React/TypeScript component updates
+- CSS/styling changes
+- Any changes in `src/flock/frontend/`
+
+#### Versioning Pattern
+
+We use **semantic versioning with beta tags**:
+- `0.5.0b56` → `0.5.0b57` for regular changes (increment beta number)
+- `0.1.2` → `0.1.3` for frontend changes (increment patch)
+
+#### How to Bump
+
+```bash
+# 1. Make your code changes
+# 2. Bump versions BEFORE committing
+# Edit pyproject.toml: version = "0.5.0b57"
+# Edit package.json: "version": "0.1.3"
+
+# 3. Commit version bumps with your changes OR as separate commit
+git add pyproject.toml src/flock/frontend/package.json
+git commit -m "chore: bump version to 0.5.0b57"
+
+# 4. Push (pre-commit hook will verify versions were bumped)
+git push
+```
+
+#### Pre-commit Hook Protection
+
+The repository has a pre-commit hook that **checks if versions were bumped**. If you forget, it will warn you:
+
+```bash
+check if version bump needed.............................................Failed
+- hook id: version-check
+
+❌ No version changes detected in modified Python/frontend files
+```
+
+#### Quick Checklist for PRs
+
+Before creating a PR, verify:
+- [ ] Backend version bumped in `pyproject.toml` (if any Python changes)
+- [ ] Frontend version bumped in `package.json` (if any UI changes)
+- [ ] Version bump committed (separate commit is fine)
+- [ ] Pre-commit hooks pass
+- [ ] **PR targets `main` branch** ✅
+
+#### ⚠️ CRITICAL: PR Base Branch
+
+**ALL pull requests MUST target the `main` branch!**
+
+```bash
+# ✅ CORRECT: PR into main branch
+gh pr create --base main --title "..." --body "..."
+
+# ❌ WRONG: PR into other branches without discussion
+gh pr create --base some-other-branch --title "..." --body "..."
+```
+
+**Why this matters:**
+- The `main` branch is the active development branch
+- All features should be validated through PR review before merging
+- Keeps the project history clean and linear
+
+**If you accidentally create a PR against the wrong branch:**
+1. Close the incorrect PR: `gh pr close <number>`
+2. Recreate against `main`: `gh pr create --base main ...`
+
+**Remember: It's better to increment versions too often than too rarely. Each meaningful change should get a version bump!**
+
+---
+
+## 🔍 Observability & Debugging with OpenTelemetry + DuckDB
+
+Flock includes **production-grade distributed tracing** that captures every operation with full input/output data—enabling both human and AI-assisted debugging.
+
+---
+
+## 🤖 FOR AI AGENTS: Tracing is Your PRIMARY Debugging Tool
+
+**⚠️ CRITICAL FOR AI AGENTS:** When debugging or testing Flock applications, **ALWAYS start with tracing**. The trace database contains the complete execution history with full input/output capture—this is your source of truth.
+
+### Why Tracing First?
+
+**Traditional debugging (logs, print statements):**
+- ❌ Requires code changes
+- ❌ Missing context (what happened before/after)
+- ❌ No timing information
+- ❌ Incomplete input/output data
+
+**Tracing-first debugging:**
+- ✅ **Zero code changes** - just query the database
+- ✅ **Complete execution history** - every operation captured
+- ✅ **Full I/O data** - exact inputs that caused issues
+- ✅ **Timing analysis** - identify bottlenecks instantly
+- ✅ **Parent-child relationships** - understand call hierarchy
+- ✅ **Correlation IDs** - track requests across agents
+
+### AI Agent Debugging Workflow (Step-by-Step)
+
+**When a user reports an issue or asks you to debug/test:**
+
+#### Step 1: Enable Tracing (if not already enabled)
+```bash
+# Check if tracing is enabled
+cat .env | grep FLOCK_AUTO_TRACE
+
+# If not enabled, add to .env:
+FLOCK_AUTO_TRACE=true
+FLOCK_TRACE_FILE=true
+
+# Or export temporarily:
+export FLOCK_AUTO_TRACE=true FLOCK_TRACE_FILE=true
+```
+
+#### Step 2: Run the Code to Capture Trace
+```bash
+# Run the problematic script/test
+uv run python examples/path/to/script.py
+
+# Or run specific test
+uv run pytest tests/test_file.py::test_name -v
+```
+
+#### Step 3: Query Trace Database for Overview
+```python
+import duckdb
+
+conn = duckdb.connect('.flock/traces.duckdb', read_only=True)
+
+# Get recent traces
+traces = conn.execute("""
+    SELECT
+        trace_id,
+        COUNT(*) as span_count,
+        MIN(start_time) as trace_start,
+        (MAX(end_time) - MIN(start_time)) / 1000000.0 as total_duration_ms,
+        SUM(CASE WHEN status_code = 'ERROR' THEN 1 ELSE 0 END) as error_count
+    FROM spans
+    GROUP BY trace_id
+    ORDER BY trace_start DESC
+    LIMIT 10
+""").fetchall()
+
+for trace in traces:
+    print(f"Trace: {trace[0][:16]}... | Spans: {trace[1]} | Duration: {trace[3]:.2f}ms | Errors: {trace[4]}")
+```
+
+#### Step 4: Analyze Specific Trace
+```python
+# Get the most recent trace (or the one with errors)
+latest_trace_id = traces[0][0]
+
+# Get execution flow with hierarchy
+flow = conn.execute("""
+    SELECT
+        span_id,
+        parent_id,
+        name,
+        service,
+        duration_ms,
+        status_code,
+        status_description,
+        json_extract(attributes, '$.correlation_id') as correlation_id
+    FROM spans
+    WHERE trace_id = ?
+    ORDER BY start_time ASC
+""", [latest_trace_id]).fetchall()
+
+# Print hierarchical execution
+for span in flow:
+    indent = '  ' if span[1] else ''  # Indent children
+    status_icon = '✅' if span[5] == 'OK' else '❌'
+    print(f"{status_icon} {indent}{span[2]} ({span[3]}) - {span[4]:.2f}ms")
+    if span[6]:  # Error description
+        print(f"   ERROR: {span[6]}")
+```
+
+#### Step 5: Examine Input/Output Data
+```python
+# Get input that caused an error
+error_details = conn.execute("""
+    SELECT
+        name,
+        status_description,
+        json_extract(attributes, '$.input.artifacts') as input_artifacts,
+        json_extract(attributes, '$.output.value') as output_value,
+        attributes
+    FROM spans
+    WHERE trace_id = ?
+    AND status_code = 'ERROR'
+""", [latest_trace_id]).fetchall()
+
+# Inspect the exact input that caused failure
+import json
+for error in error_details:
+    print(f"\n❌ ERROR in {error[0]}")
+    print(f"Message: {error[1]}")
+    print(f"Input: {error[2]}")
+    print(f"Output: {error[3]}")
+```
+
+#### Step 6: Identify Root Cause
+```python
+# Common root cause queries:
+
+# 1. Find the slowest operation in the trace
+slowest = conn.execute("""
+    SELECT name, service, duration_ms
+    FROM spans
+    WHERE trace_id = ?
+    ORDER BY duration_ms DESC
+    LIMIT 1
+""", [latest_trace_id]).fetchone()
+print(f"Bottleneck: {slowest[0]} ({slowest[1]}) took {slowest[2]:.2f}ms")
+
+# 2. Check if agent was triggered correctly
+agent_triggers = conn.execute("""
+    SELECT
+        name,
+        json_extract(attributes, '$.input.artifacts') as consumed_artifacts
+    FROM spans
+    WHERE trace_id = ?
+    AND name LIKE 'Agent.execute'
+""", [latest_trace_id]).fetchall()
+
+# 3. Verify artifact types produced
+artifacts_produced = conn.execute("""
+    SELECT DISTINCT
+        service as agent,
+        json_extract(attributes, '$.output.type') as artifact_type
+    FROM spans
+    WHERE trace_id = ?
+    AND attributes->>'output.type' IS NOT NULL
+""", [latest_trace_id]).fetchall()
+```
+
+#### Step 7: Report Findings & Fix
+```python
+# Close connection
+conn.close()
+
+# Now you have:
+# - Exact execution flow
+# - Input data that caused the issue
+# - Timing information (bottlenecks)
+# - Error messages and stack traces
+# - Artifact flow between agents
+
+# Report to user with specific findings
+print("""
+DIAGNOSIS COMPLETE:
+
+Issue: <describe the problem>
+Root Cause: <specific operation/input that failed>
+Evidence:
+  - Trace ID: {trace_id}
+  - Failed at: {operation_name}
+  - Input: {input_data}
+  - Duration: {duration}ms
+
+Recommendation: <how to fix>
+""")
+```
+
+### Essential Queries for AI Agents
+
+**Keep these queries ready for common debugging tasks:**
+
+#### 1. Find Most Recent Workflow Execution
+```python
+latest_workflow = conn.execute("""
+    SELECT trace_id,
+           COUNT(*) as operations,
+           (MAX(end_time) - MIN(start_time)) / 1000000.0 as duration_ms
+    FROM spans
+    GROUP BY trace_id
+    ORDER BY MIN(start_time) DESC
+    LIMIT 1
+""").fetchone()
+```
+
+#### 2. Check Agent Lifecycle Execution
+```python
+# Verify all lifecycle hooks fired correctly
+lifecycle = conn.execute("""
+    SELECT name, duration_ms, status_code
+    FROM spans
+    WHERE trace_id = ?
+    AND service LIKE '%Component'
+    OR service LIKE '%Engine'
+    ORDER BY start_time ASC
+""", [trace_id]).fetchall()
+
+# Expected order: on_initialize → on_pre_consume → on_pre_evaluate →
+#                 evaluate → on_post_evaluate → on_post_publish → on_terminate
+```
+
+#### 3. Validate Artifact Flow
+```python
+# Track artifact transformations
+artifact_flow = conn.execute("""
+    SELECT
+        name,
+        service,
+        json_extract(attributes, '$.input.artifacts[0].type') as input_type,
+        json_extract(attributes, '$.output.type') as output_type
+    FROM spans
+    WHERE trace_id = ?
+    AND (attributes->>'input.artifacts' IS NOT NULL
+         OR attributes->>'output.type' IS NOT NULL)
+    ORDER BY start_time ASC
+""", [trace_id]).fetchall()
+
+# Verify expected transformations: InputType → Agent → OutputType
+```
+
+#### 4. Detect Performance Issues
+```python
+# Find operations that took >1 second
+slow_ops = conn.execute("""
+    SELECT
+        name,
+        service,
+        duration_ms,
+        json_extract(attributes, '$.input.artifacts[0].payload') as input_payload
+    FROM spans
+    WHERE trace_id = ?
+    AND duration_ms > 1000
+    ORDER BY duration_ms DESC
+""", [trace_id]).fetchall()
+
+# Check if large payloads are causing slowness
+for op in slow_ops:
+    if op[3]:
+        payload_size = len(str(op[3]))
+        print(f"{op[0]}: {op[2]:.0f}ms (payload: {payload_size} bytes)")
+```
+
+#### 5. Debug Test Failures
+```python
+# When a test fails, find what actually happened vs expected
+test_execution = conn.execute("""
+    SELECT
+        name,
+        status_code,
+        status_description,
+        json_extract(attributes, '$.input.artifacts') as input,
+        json_extract(attributes, '$.output.value') as output
+    FROM spans
+    WHERE trace_id = ?
+    ORDER BY start_time ASC
+""", [trace_id]).fetchall()
+
+# Compare actual output with test expectations
+```
+
+### Common Debugging Scenarios for AI Agents
+
+#### Scenario A: "Test is failing but I don't know why"
+```bash
+# Step 1: Run test with tracing
+FLOCK_AUTO_TRACE=true FLOCK_TRACE_FILE=true uv run pytest tests/test_file.py::test_name -v
+
+# Step 2: Query for test execution
+uv run python -c "
+import duckdb
+conn = duckdb.connect('.flock/traces.duckdb', read_only=True)
+
+# Find most recent trace
+trace = conn.execute('''
+    SELECT trace_id FROM spans
+    GROUP BY trace_id
+    ORDER BY MIN(start_time) DESC LIMIT 1
+''').fetchone()[0]
+
+# Get all operations
+ops = conn.execute('''
+    SELECT name, status_code, duration_ms
+    FROM spans WHERE trace_id = ?
+    ORDER BY start_time
+''', [trace]).fetchall()
+
+for op in ops:
+    status = '✅' if op[1] == 'OK' else '❌'
+    print(f'{status} {op[0]}: {op[2]:.2f}ms')
+"
+```
+
+#### Scenario B: "Agent not producing expected output"
+```python
+import duckdb
+conn = duckdb.connect('.flock/traces.duckdb', read_only=True)
+
+# Find what the agent actually produced
+trace_id = '<latest_trace_id>'
+output = conn.execute("""
+    SELECT
+        service as agent_name,
+        json_extract(attributes, '$.output.type') as output_type,
+        json_extract(attributes, '$.output.value') as output_value
+    FROM spans
+    WHERE trace_id = ?
+    AND name = 'Agent.execute'
+""", [trace_id]).fetchone()
+
+print(f"Agent: {output[0]}")
+print(f"Output Type: {output[1]}")
+print(f"Output Value: {output[2]}")
+
+# Compare with expected output type in test
+```
+
+#### Scenario C: "Agent not being triggered"
+```python
+# Check if artifact was published and if agent subscribed
+trace_id = '<latest_trace_id>'
+
+published = conn.execute("""
+    SELECT json_extract(attributes, '$.output.type') as artifact_type
+    FROM spans
+    WHERE trace_id = ?
+    AND name = 'Flock.publish'
+""", [trace_id]).fetchone()
+
+print(f"Published artifact type: {published[0]}")
+
+# Check if any agent consumed it
+consumers = conn.execute("""
+    SELECT service, json_extract(attributes, '$.input.artifacts[0].type') as consumed_type
+    FROM spans
+    WHERE trace_id = ?
+    AND name = 'Agent.execute'
+""", [trace_id]).fetchall()
+
+if not consumers:
+    print("❌ No agents consumed this artifact!")
+    print("Check agent subscription rules (consumes clause)")
+else:
+    for consumer in consumers:
+        print(f"✅ {consumer[0]} consumed {consumer[1]}")
+```
+
+#### Scenario D: "Performance regression"
+```python
+# Compare execution times across traces
+import duckdb
+conn = duckdb.connect('.flock/traces.duckdb', read_only=True)
+
+# Get last 10 executions of same operation
+perf_history = conn.execute("""
+    SELECT
+        trace_id,
+        start_time,
+        duration_ms
+    FROM spans
+    WHERE name = 'DSPyEngine.evaluate'  -- or any operation
+    ORDER BY start_time DESC
+    LIMIT 10
+""").fetchall()
+
+# Calculate average and detect outliers
+durations = [p[2] for p in perf_history]
+avg = sum(durations) / len(durations)
+latest = durations[0]
+
+print(f"Latest: {latest:.2f}ms")
+print(f"Average: {avg:.2f}ms")
+print(f"Change: {((latest / avg - 1) * 100):+.1f}%")
+
+if latest > avg * 1.5:
+    print("⚠️ Performance regression detected!")
+```
+
+### Best Practices for AI Agents
+
+**✅ DO:**
+- **Always enable tracing** before running code to debug
+- **Start with overview queries** (get all traces, find latest)
+- **Work from general to specific** (trace → spans → attributes → I/O)
+- **Use read-only connections** (`read_only=True`)
+- **Close connections** when done
+- **Clear old traces** periodically: `Flock.clear_traces()`
+- **Use correlation IDs** to track related operations
+
+**❌ DON'T:**
+- Don't modify code just to add logging - query traces instead
+- Don't guess at execution flow - trace database has the truth
+- Don't skip trace analysis for "obvious" bugs - verify with data
+- Don't forget to check `status_description` for error details
+- Don't ignore timing data - it reveals bottlenecks
+
+### Tracing-First Testing Workflow
+
+When writing or debugging tests:
+
+1. **Run test with tracing enabled**
+   ```bash
+   FLOCK_AUTO_TRACE=true FLOCK_TRACE_FILE=true uv run pytest tests/test_file.py -v
+   ```
+
+2. **Query trace to see what actually happened**
+   ```python
+   # Get test execution trace
+   trace_id = conn.execute("SELECT trace_id FROM spans GROUP BY trace_id ORDER BY MIN(start_time) DESC LIMIT 1").fetchone()[0]
+   ```
+
+3. **Verify assertions match reality**
+   ```python
+   # What did the agent actually produce?
+   actual = conn.execute("SELECT json_extract(attributes, '$.output.value') FROM spans WHERE trace_id = ? AND name = 'Agent.execute'", [trace_id]).fetchone()
+
+   # Does it match test expectations?
+   expected = "BugDiagnosis artifact with severity='Critical'"
+   ```
+
+4. **Debug failures with exact I/O data**
+   ```python
+   # Get the exact input that caused test failure
+   failure_input = conn.execute("""
+       SELECT json_extract(attributes, '$.input.artifacts')
+       FROM spans WHERE trace_id = ? AND status_code = 'ERROR'
+   """, [trace_id]).fetchone()
+   ```
+
+### Quick Start: Enable Tracing
+
+```bash
+# Enable auto-tracing (add to .env or export)
+FLOCK_AUTO_TRACE=true      # Enable tracing for all operations
+FLOCK_TRACE_FILE=true      # Store traces in DuckDB
+
+# Run your application
+python examples/showcase/01_declarative_pizza.py
+
+# Traces stored in: .flock/traces.duckdb
+```
+
+### 🆕 Unified Tracing with traced_run()
+
+**New in v0.5.0**: Wrap entire workflows in a single parent trace for better observability!
+
+```python
+# ✅ Unified trace
+async with flock.traced_run("pizza_workflow"):
+    await flock.publish(pizza_idea)
+    await flock.run_until_idle()
+```
+
+**Result**: Single trace with proper hierarchy
+
+### 🗑️ Clearing Traces
+
+Clear trace database for fresh debug sessions:
+
+```python
+# Clear all traces
+result = Flock.clear_traces()
+print(f"Deleted {result['deleted_count']} spans")
+```
+
+📖 **Full tracing documentation**: [docs/UNIFIED_TRACING.md](docs/UNIFIED_TRACING.md)
+
+---
+
+## ❓ FAQ for AI Agents
+
+### Q: How do I debug or test Flock code?
+
+**⚠️ ALWAYS START WITH TRACING!**
+
+Quick debugging workflow:
+```bash
+# 1. Enable tracing
+export FLOCK_AUTO_TRACE=true FLOCK_TRACE_FILE=true
+
+# 2. Run the code
+uv run python examples/path/to/script.py
+
+# 3. Query trace database
+uv run python -c "
+import duckdb
+conn = duckdb.connect('.flock/traces.duckdb', read_only=True)
+
+# Get latest trace
+trace = conn.execute('''
+    SELECT trace_id, COUNT(*) as spans,
+           (MAX(end_time)-MIN(start_time))/1000000.0 as duration_ms
+    FROM spans GROUP BY trace_id
+    ORDER BY MIN(start_time) DESC LIMIT 1
+''').fetchone()
+
+print(f'Trace: {trace[0][:32]}...')
+print(f'Spans: {trace[1]}, Duration: {trace[2]:.2f}ms')
+
+# Get execution flow
+flow = conn.execute('''
+    SELECT name, duration_ms, status_code
+    FROM spans WHERE trace_id = ?
+    ORDER BY start_time
+''', [trace[0]]).fetchall()
+
+for op in flow:
+    status = '✅' if op[2] == 'OK' else '❌'
+    print(f'{status} {op[0]}: {op[1]:.2f}ms')
+"
+```
+
+📖 **Complete guide:** [🤖 FOR AI AGENTS: Tracing is Your PRIMARY Debugging Tool](#-for-ai-agents-tracing-is-your-primary-debugging-tool) (scroll up in this file)
+
+### Q: Where should I save new files?
+
+**⚠️ CRITICAL: NEVER save files to the project root directory!**
+
+Always use the appropriate subdirectory:
+- **Tests**: `/tests` - All test files only
+- **Source Code**: `/src/flock` - Production code following refactored module structure:
+  - `/src/flock/core/` - Core orchestration and agents (Phase 2)
+  - `/src/flock/orchestrator/` - Orchestrator modules (Phase 3 + 5A)
+  - `/src/flock/agent/` - Agent modules (Phase 4)
+  - `/src/flock/components/` - Component library
+  - `/src/flock/engines/` - Engine implementations
+  - `/src/flock/utils/` - Utility modules (Phase 1)
+- **Documentation**: `/docs` - Documentation only
+  - `/docs/patterns/` - Code patterns and best practices
+  - `/docs/guides/` - User guides and tutorials
+  - `/docs/refactor/` - Refactoring documentation (Phase 1-7)
+- **Examples**: `/examples` - Example scripts only
+- **Frontend**: `src/flock/frontend/src` - React components and frontend code
+
+**Never create files in the root directory** - it should only contain configuration files like `pyproject.toml`, `README.md`, etc.
+
+**📐 See [Architecture Overview](docs/architecture.md) for complete module organization details.**
+
+### Q: How do I add a new dependency?
+
+```bash
+# Python
+uv add package-name
+
+# Frontend
+cd frontend && npm install package-name
+```
+
+📖 **Full guide:** [Dependencies Guide](docs/ai-agents/dependencies.md)
+
+### Q: How do I run a specific test?
+
+```bash
+# Backend
+uv run pytest tests/test_file.py::test_name -v
+
+# Frontend
+cd frontend && npm test -- test_name
+```
+
+### Q: How do I start the dashboard?
+
+```python
+await orchestrator.serve(dashboard=True)
+```
+
+📖 **Full guide:** [Frontend Guide](docs/ai-agents/frontend.md)
+
+### Q: How do I test UI features and debug dashboard issues?
+
+**Use playwright-mcp with dashboard examples for comprehensive manual UI testing.**
+
+#### Step-by-Step Testing Workflow
+
+**1. Start the Dashboard Example**
+```bash
+# Run in background to keep testing
+uv run python examples/03-the-dashboard/01_declarative_pizza.py
+```
+
+Wait for these success indicators in the output:
+- `[Dashboard] Production build completed`
+- `INFO: Uvicorn running on http://127.0.0.1:8344`
+- `[Dashboard] Browser launched successfully`
+
+**2. Navigate and Verify Initial Load**
+```python
+# Use playwright-mcp tools
+mcp__playwright__browser_navigate(url="http://localhost:8344")
+```
+
+**✅ What to verify:**
+- Page title: "🦆🐓 Flock 🐤🐧"
+- WebSocket status (top right): **"Connected"** (green)
+- Two view buttons: "Agent View" (active) and "Blackboard View"
+- Control buttons: Publish, Agent Details, Filters, Settings
+
+**3. Test Agent View**
+
+Take a snapshot to see the agent graph:
+```python
+mcp__playwright__browser_snapshot()
+```
+
+**✅ What to verify in Agent View:**
+- Agent nodes displayed (e.g., `pizza_master`)
+- Each node shows:
+  - Agent name and status (should be "idle" initially)
+  - Input types with count (e.g., ↓ 0 __main__.MyDreamPizza)
+  - Output types with count (e.g., ↑ 0 __main__.Pizza)
+- React Flow controls (zoom in/out, fit view, mini-map)
+
+**4. Open Agent Details Panel**
+```python
+# Click the Agent Details button
+mcp__playwright__browser_click(element="Agent Details button", ref="<ref>")
+```
+
+**✅ What to verify:**
+- Panel opens showing agent name
+- Three tabs: "Live Output", "Message History", "Run Status"
+- Shows "Idle - no output" initially
+
+**5. Test Publishing an Artifact**
+
+```python
+# Step 1: Select artifact type
+mcp__playwright__browser_select_option(
+    element="Artifact Type dropdown",
+    ref="<ref>",
+    values=["__main__.MyDreamPizza"]
+)
+```
+
+**✅ What to verify after selecting type:**
+- Form dynamically generates input fields based on artifact schema
+- For MyDreamPizza: Should show "Pizza Idea" textbox
+
+```python
+# Step 2: Fill in the input
+mcp__playwright__browser_type(
+    element="Pizza Idea textbox",
+    ref="<ref>",
+    text="a spicy Hawaiian pizza with jalapeños and pineapple"
+)
+
+# Step 3: Publish
+mcp__playwright__browser_click(
+    element="Publish Artifact button",
+    ref="<ref>"
+)
+```
+
+**✅ What to verify after publishing:**
+- Agent status changes: "idle" → "running" → "idle"
+- Input count increases: ↓ 0 → ↓ 1
+- Output count increases: ↑ 0 → ↑ 1
+- **External node appears** on the graph (shows who published the artifact)
+- Edge connects external → pizza_master
+
+**6. Monitor Live Execution**
+
+**✅ What to verify in Agent Details panel:**
+- Event counter increases (e.g., "316 events")
+- Live streaming output appears token-by-token
+- Console shows WebSocket messages: `[WebSocket] Streaming output: {...}`
+- Final output shows complete structured data
+- "--- End of output ---" marker when complete
+
+**7. Test Blackboard View**
+```python
+# Switch to Blackboard View
+mcp__playwright__browser_click(
+    element="Blackboard View button",
+    ref="<ref>"
+)
+```
+
+**✅ What to verify in Blackboard View:**
+- **Input artifact node** showing:
+  - Type: `__main__.MyDreamPizza`
+  - Producer: "by: external"
+  - Timestamp
+  - Full JSON payload (expandable)
+- **Output artifact node** showing:
+  - Type: `__main__.Pizza`
+  - Producer: "by: pizza_master"
+  - Timestamp
+  - Complete structured data with all fields
+- **Edge** connecting input → output artifacts
+- Data is fully browsable (can expand/collapse nested objects)
+
+**8. Take Screenshots for Verification**
+```python
+# Capture key states for visual verification
+mcp__playwright__browser_take_screenshot(filename="dashboard-test.png")
+```
+
+#### Common Issues and Troubleshooting
+
+**Issue: Dashboard doesn't load**
+- Check: Backend server started? Look for "Uvicorn running" message
+- Check: Frontend build completed? Look for "Production build completed"
+- Solution: Wait 5-10 seconds after starting for build to complete
+
+**Issue: WebSocket shows "Disconnected"**
+- Check: Console for WebSocket errors
+- Check: Server logs for WebSocket connection messages
+- Solution: Refresh page, verify server is running
+
+**Issue: No live output during agent execution**
+- Check: Agent Details panel is open
+- Check: "Live Output" tab is active
+- Check: Console shows `[WebSocket] Streaming output` messages
+- Solution: Verify WebSocket connection status
+
+**Issue: Artifacts not appearing in Blackboard View**
+- Check: Did agent execution complete? (status back to "idle")
+- Check: Output count increased? (↑ 1)
+- Solution: Switch back to Agent View to verify execution, then return to Blackboard View
+
+#### Why Manual Testing with Playwright-MCP?
+
+- ✅ **Live testing** - Real dashboard with actual agents executing
+- ✅ **Visual verification** - See exactly what users see
+- ✅ **WebSocket testing** - Verify real-time streaming works correctly
+- ✅ **Full workflow** - Test complete user journey from publish → execute → view results
+- ✅ **Screenshot capture** - Document UI state for debugging/documentation
+- ✅ **Interactive debugging** - Click, type, inspect like a real user
+
+📖 **Dashboard examples:** [`examples/02-dashboard/`](examples/02-dashboard/)
+
+---
+
+#### Advanced Dashboard Testing: Multi-Agent Cascades & Conditional Consumption
+
+**Test with:** `examples/03-the-dashboard/02-dashboard-edge-cases.py`
+
+This example demonstrates advanced features not visible in simple single-agent workflows.
+
+**1. Use Auto Layout for Complex Graphs**
+
+When testing with 3+ agents, use the context menu to organize the graph:
+
+```python
+# After navigating to dashboard
+# Right-click on the canvas (not on a node)
+mcp__playwright__browser_click(element="Canvas area", ref="<ref>", button="right")
+
+# Click Auto Layout from context menu
+mcp__playwright__browser_click(element="Auto Layout button", ref="<ref>")
+```
+
+**✅ What to verify:**
+- Agents arranged in clean vertical or horizontal hierarchy
+- No overlapping nodes
+- Edges clearly visible between agents
+
+**2. Test Conditional Consumption (Lambda Filters)**
+
+This example has agents with `where` clauses that filter which artifacts they consume:
+
+```python
+# In the code:
+chapter_agent.consumes(Review, where=lambda r: r.score >= 9)  # Only high scores
+book_idea_agent.consumes(Review, where=lambda r: r.score <= 8)  # Only low scores (feedback loop)
+```
+
+**✅ What to verify after publishing Idea:**
+- **Edge labels show filtered counts**: e.g., `__main__.Review(1)` means "1 Review consumed out of 3 total"
+- **Input counts reflect actual consumption**: chapter_agent shows ↓ 1 Review (not ↓ 3)
+- **Feedback loops work**: book_idea_agent consumes both Idea AND Review artifacts
+
+**3. Monitor Multi-Agent Cascade Execution**
+
+Expected workflow for edge cases example:
+```
+1. Publish Idea artifact
+2. book_idea_agent: ↓ 1 Idea → ↑ 3 BookHook (produces multiple outputs!)
+3. reviewer_agent: ↓ 3 BookHook → ↑ 3 Review (processes each hook)
+4. chapter_agent: ↓ 1 Review → ↑ 1 BookOutline (filtered: only score >= 9)
+5. book_idea_agent: ↓ 2 Review → ↑ 0 BookHook (feedback loop for low scores)
+```
+
+**✅ What to verify during cascade:**
+- Agent statuses transition: idle → running → idle
+- Counters update in real-time as each agent completes
+- External node persists showing initial publisher
+- Edges appear/update showing data flow
+
+**4. Handle Large Artifact Counts**
+
+**⚠️ IMPORTANT**: When page has 8+ artifacts, `browser_snapshot()` exceeds 25K token limit and fails.
+
+**Solution**: Use `browser_take_screenshot()` for visual verification instead:
+
+```python
+# ❌ This will fail with many artifacts
+mcp__playwright__browser_snapshot()
+
+# ✅ Use screenshots instead
+mcp__playwright__browser_take_screenshot(filename="cascade-state.png")
+```
+
+**5. Verify Final State in Blackboard View**
+
+```python
+# Switch to Blackboard View
+mcp__playwright__browser_click(element="Blackboard View button", ref="<ref>")
+
+# Take screenshot (snapshot will fail with many artifacts)
+mcp__playwright__browser_take_screenshot(filename="blackboard-artifacts.png")
+```
+
+**✅ What to verify:**
+- All produced artifacts visible as nodes
+- **1 Idea** → **3 BookHooks** → **3 Reviews** → **1 BookOutline**
+- Edges show complete transformation chain
+- Final artifact (BookOutline) contains expected structured data
+- Timestamps show execution order
+
+#### Expected Execution Time
+
+**Simple example (01_declarative_pizza.py)**: ~5 seconds (1 agent, 1 artifact)
+**Edge cases (02-dashboard-edge-cases.py)**: ~60 seconds (3 agents, 8 artifacts, feedback loop)
+
+Plan your testing time accordingly!
+
+#### Key Dashboard Features Learned
+
+**Auto Layout** ⭐
+- Access via right-click context menu
+- Automatically organizes complex agent graphs
+- Essential for 3+ agent workflows
+
+**Filtered Edge Labels** ⭐
+- Shows actual consumed count vs total available
+- Format: `ArtifactType(consumed_count)`
+- Makes conditional consumption transparent
+
+**Feedback Loops** ⭐
+- Agents can consume multiple artifact types
+- Low-scoring Reviews loop back to book_idea_agent
+- Counts accumulate correctly across iterations
+
+**Real-time Updates** ⭐
+- Status changes: idle → running → idle
+- Counters increment as artifacts produced/consumed
+- WebSocket delivers updates without page refresh
+
+📖 **Dashboard examples:** [`examples/02-dashboard/`](examples/02-dashboard/)
+
+---
+
+### Q: How do I use unified tracing?
+
+```python
+# Wrap workflows in a single trace
+async with flock.traced_run("workflow_name"):
+    await flock.publish(data)
+    await flock.run_until_idle()
+
+# Clear traces for fresh debug session
+result = Flock.clear_traces()
+```
+
+📖 **Full guide:** [docs/UNIFIED_TRACING.md](docs/UNIFIED_TRACING.md)
+
+### Q: Where should I add new tests?
+
+Add to existing test file if relevant, or create new file following naming convention `test_<module>.py` for backend, `<name>.test.tsx` for frontend.
+
+### Q: What Python version features can I use?
+
+Python 3.12+, so you can use:
+- `match`/`case` statements
+- `TaskGroup` for parallel execution (see [docs/patterns/async_patterns.md](docs/patterns/async_patterns.md))
+- Improved type hints (`list[str]` not `List[str]`)
+- Exception groups with `except*` (see [docs/patterns/error_handling.md](docs/patterns/error_handling.md))
+
+**⚠️ CRITICAL:** Follow error handling and async patterns documented in `/docs/patterns/` - these are REQUIRED for all new code.
+
+### Q: How do I debug WebSocket issues?
+
+Check browser console for WebSocket logs, use Network tab to inspect connection, and verify backend WebSocket server is running on correct port.
+
+---
+
+## 🎯 Quick Reference
+
+### Essential Commands
+
+```bash
+# Setup
+poe install          # Install all dependencies
+
+# Development
+poe build           # Build project
+poe lint            # Lint code
+poe format          # Format code
+
+# Testing
+poe test            # Run tests
+poe test-cov        # Run with coverage
+poe test-critical   # Run critical path tests
+
+# Frontend
+cd frontend
+npm run dev         # Start dev server
+npm test            # Run frontend tests
+npm run build       # Build for production
+
+# UI Testing (with playwright-mcp)
+uv run python examples/03-the-dashboard/01_declarative_pizza.py  # Start dashboard
+# Then use playwright-mcp to interact with UI for manual testing
+```
+
+### Code Snippets
+
+**Create orchestrator:**
+```python
+from flock import Flock  # ✅ Correct import (refactored in Phase 3)
+orchestrator = Flock("openai/gpt-4o")
+```
+
+**Define agent:**
+```python
+agent = (
+    orchestrator.agent("name")
+    .description("What it does")
+    .consumes(InputType)
+    .publishes(OutputType)
+)
+```
+
+**Run agent:**
+```python
+await orchestrator.arun(agent, input_data)
+```
+
+**Fan-out publishing (multiple outputs):**
+```python
+# Single-type fan-out: Generate 10 outputs per execution
+agent = (
+    orchestrator.agent("generator")
+    .consumes(InputType)
+    .publishes(OutputType, fan_out=10)
+)
+
+# Multi-output fan-out: Generate 3 of EACH type = 9 total artifacts in ONE LLM call!
+multi_master = (
+    orchestrator.agent("multi_master")
+    .consumes(Idea)
+    .publishes(Movie, MovieScript, MovieCampaign, fan_out=3)
+)
+# Result: 3 Movies + 3 MovieScripts + 3 MovieCampaigns = 9 artifacts
+# 89% cost savings vs 9 separate calls + perfect context alignment!
+
+# With filtering (reduce noise)
+agent.publishes(OutputType, fan_out=20, where=lambda o: o.score >= 8.0)
+
+# With validation (enforce quality)
+agent.publishes(
+    OutputType,
+    fan_out=10,
+    validate=lambda o: o.field in ["valid", "values"]
+)
+
+# With dynamic visibility (per-artifact access control)
+agent.publishes(
+    OutputType,
+    fan_out=5,
+    visibility=lambda o: PrivateVisibility(agents=[o.recipient])
+)
+```
+
+**Unified tracing:**
+```python
+# Wrap workflow in single trace
+async with flock.traced_run("workflow_name"):
+    await flock.publish(data)
+    await flock.run_until_idle()
+
+# Clear traces
+Flock.clear_traces()
+```
+
+**Start dashboard:**
+```python
+await orchestrator.serve(dashboard=True)
+```
+
+---
+
+## 📞 Getting Help
+
+### Documentation
+
+**AI Agent Guides (this repo):**
+- **[Architecture Overview](docs/architecture.md)** - Complete system design, module organization, Phase 1-7 refactoring ⭐ **MUST READ**
+- **[Error Handling Patterns](docs/patterns/error_handling.md)** - Production error handling guide ⭐ **REQUIRED**
+- **[Async Patterns](docs/patterns/async_patterns.md)** - Async/await best practices ⭐ **REQUIRED**
+- **[Architecture Guide](docs/ai-agents/architecture.md)** - Core architecture, project structure, code style
+- **[Development Workflow](docs/ai-agents/development.md)** - Testing, quality standards, versioning
+- **[Frontend Guide](docs/ai-agents/frontend.md)** - Dashboard usage, frontend development
+- **[Dependencies Guide](docs/ai-agents/dependencies.md)** - Package management, UV commands
+- **[Common Tasks](docs/ai-agents/common-tasks.md)** - Adding agents/components, performance
+
+**Additional Documentation:**
+- **Contributing Guide:** [`CONTRIBUTING.md`](CONTRIBUTING.md) - Complete contribution workflow
+- **Versioning Guide:** [`docs/VERSIONING.md`](docs/VERSIONING.md) - Smart version bumping
+- **Pre-commit Hooks:** [`docs/PRE_COMMIT_HOOKS.md`](docs/PRE_COMMIT_HOOKS.md) - Quality automation
+- **Unified Tracing:** [`docs/UNIFIED_TRACING.md`](docs/UNIFIED_TRACING.md) - Workflow tracing & trace management
+- **Examples:** [`examples/`](examples/) - Working code examples
+- **Analysis Documents:** [`docs/patterns/`](docs/patterns/)
+
+---
+
+**Welcome to the team! Let's build the future of AI agent orchestration together.** 🚀
+
+---
+
+*Last updated: October 19, 2025*
+*This file follows the modern AGENTS.md format for AI coding agents.*
+*Reflects Phase 1-7 refactoring with updated module organization and patterns.*