docs: add context management deep dive

docs/guides/configuration-reference.md

@@ -1,453 +0,0 @@
-# Configuration Reference Guide
-
-This guide provides a comprehensive reference for configuring CleverAgents. Configuration is primarily managed through environment variables, with sensible defaults for most settings.
-
-## Overview
-
-CleverAgents uses a **Pydantic-based settings system** that reads configuration from environment variables with the `CLEVERAGENTS_` prefix. The configuration system is designed to be:
-
-- **Environment-first**: All configuration comes from environment variables, making it ideal for containerized and cloud-native deployments
-- **Hierarchical**: Settings are organized by functional area (runtime, providers, storage, observability, etc.)
-- **Validated**: All configuration values are validated at startup with clear error messages
-- **Sensible defaults**: Most settings have reasonable defaults suitable for development and production
-
-### Configuration Loading Order
-
-1. **Environment variables** (highest priority) — `CLEVERAGENTS_*` prefixed variables
-2. **Provider-specific environment variables** — Native provider variables (e.g., `OPENAI_API_KEY`)
-3. **Default values** — Hardcoded defaults in the settings schema
-
-When a setting is not explicitly configured, the system falls back to the next level in the order above.
-
-### Configuration File Locations
-
-CleverAgents stores configuration and data in the following locations:
-
-| Purpose | Default Location | Environment Variable |
-|---------|------------------|----------------------|
-| User data (sessions, personas, etc.) | `~/.cleveragents/` | `CLEVERAGENTS_DATA_DIR` |
-| Database | `~/.cleveragents/cleveragents.db` | `CLEVERAGENTS_DATABASE_URL` |
-| Logs | `./logs/` | `CLEVERAGENTS_LOG_DIR` |
-| Storage (plans, resources, etc.) | `./data/` | `CLEVERAGENTS_STORAGE_BASE_PATH` |
-| Vector store | `./.cleveragents/vector_store/` | `CLEVERAGENTS_VECTOR_STORE_PATH` |
-
-## Environment Variables Reference
-
-All environment variables are organized by functional area. Each entry includes:
-
-- **Variable name** — The full `CLEVERAGENTS_*` environment variable name
-- **Type** — The expected data type (string, integer, boolean, etc.)
-- **Default** — The default value if not set
-- **Description** — What the setting controls
-- **Example** — A practical example value
-
-### Runtime & Server Configuration
-
-#### `CLEVERAGENTS_ENV`
-- **Type:** String
-- **Default:** `development`
-- **Valid values:** `development`, `production`, `staging`, `test`
-- **Description:** The runtime environment. Controls logging verbosity, debug features, and validation strictness.
-- **Example:** `export CLEVERAGENTS_ENV=production`
-
-#### `CLEVERAGENTS_DEBUG_ENABLED`
-- **Type:** Boolean
-- **Default:** `false`
-- **Description:** Enable debug mode with verbose logging and additional runtime checks. Should never be enabled in production.
-- **Example:** `export CLEVERAGENTS_DEBUG_ENABLED=true`
-
-#### `CLEVERAGENTS_SERVER_HOST`
-- **Type:** String (IP address)
-- **Default:** `0.0.0.0`
-- **Description:** The host address for the CleverAgents server to bind to.
-- **Example:** `export CLEVERAGENTS_SERVER_HOST=127.0.0.1`
-
-#### `CLEVERAGENTS_SERVER_PORT`
-- **Type:** Integer
-- **Default:** `8080`
-- **Valid range:** 1-65535
-- **Description:** The port for the CleverAgents server to listen on.
-- **Example:** `export CLEVERAGENTS_SERVER_PORT=9000`
-
-#### `CLEVERAGENTS_SERVER_RELOAD`
-- **Type:** Boolean
-- **Default:** `false`
-- **Description:** Enable auto-reload on code changes (development only). Never use in production.
-- **Example:** `export CLEVERAGENTS_SERVER_RELOAD=true`
-
-#### `CLEVERAGENTS_SERVER_URL`
-- **Type:** String (URL)
-- **Default:** `None` (local mode)
-- **Description:** URL of a remote CleverAgents server to connect to. When set, the CLI operates in server mode.
-- **Example:** `export CLEVERAGENTS_SERVER_URL=https://cleveragents.example.com`
-
-#### `CLEVERAGENTS_SERVER_TOKEN`
-- **Type:** String
-- **Default:** `None`
-- **Description:** Authentication token for connecting to a remote CleverAgents server. Required when `CLEVERAGENTS_SERVER_URL` is set.
-- **Example:** `export CLEVERAGENTS_SERVER_TOKEN=sk-1234567890abcdef`
-
-### AI Provider Configuration
-
-#### Provider API Keys
-
-Each AI provider requires its API key to be configured. CleverAgents supports multiple providers simultaneously and automatically selects the best available one.
-
-| Provider | Primary Variable | Fallback Variables | Example |
-|----------|------------------|-------------------|---------|
-| OpenAI | `OPENAI_API_KEY` | `CLEVERAGENTS_OPENAI_API_KEY` | `sk-proj-...` |
-| Anthropic | `ANTHROPIC_API_KEY` | `CLEVERAGENTS_ANTHROPIC_API_KEY` | `sk-ant-...` |
-| Google | `GOOGLE_API_KEY` | `GOOGLE_GENAI_API_KEY`, `CLEVERAGENTS_GOOGLE_API_KEY` | `AIza...` |
-| Azure OpenAI | `AZURE_OPENAI_API_KEY` | `AZURE_API_KEY`, `CLEVERAGENTS_AZURE_API_KEY` | `...` |
-| OpenRouter | `OPENROUTER_API_KEY` | `CLEVERAGENTS_OPENROUTER_API_KEY` | `sk-or-...` |
-| Gemini | `GEMINI_API_KEY` | `GOOGLE_GEMINI_API_KEY`, `CLEVERAGENTS_GEMINI_API_KEY` | `AIza...` |
-| Cohere | `COHERE_API_KEY` | `CLEVERAGENTS_COHERE_API_KEY` | `...` |
-| Groq | `GROQ_API_KEY` | `CLEVERAGENTS_GROQ_API_KEY` | `gsk_...` |
-| Together | `TOGETHER_API_KEY` | `CLEVERAGENTS_TOGETHER_API_KEY` | `...` |
-| HuggingFace | `HF_TOKEN` | `HUGGINGFACEHUB_API_TOKEN`, `HUGGING_FACE_HUB_TOKEN` | `hf_...` |
-| Perplexity | `PERPLEXITY_API_KEY` | `CLEVERAGENTS_PERPLEXITY_API_KEY` | `pplx-...` |
-
-#### `CLEVERAGENTS_DEFAULT_PROVIDER`
-- **Type:** String
-- **Default:** `None` (auto-detect from configured providers)
-- **Valid values:** `openai`, `anthropic`, `google`, `azure`, `openrouter`, `gemini`, `cohere`, `groq`, `together`, `huggingface`, `perplexity`
-- **Description:** Pin the default AI provider. When not set, the system automatically selects the first available configured provider in the fallback order.
-- **Example:** `export CLEVERAGENTS_DEFAULT_PROVIDER=anthropic`
-
-#### `CLEVERAGENTS_DEFAULT_MODEL`
-- **Type:** String
-- **Default:** `None` (provider-specific default)
-- **Description:** Pin the default model for the selected provider. When not set, each provider's published default is used (e.g., `gpt-4o` for OpenAI, `claude-sonnet-4-20250514` for Anthropic).
-- **Example:** `export CLEVERAGENTS_DEFAULT_MODEL=gpt-4-turbo`
-
-#### `CLEVERAGENTS_MOCK_PROVIDERS`
-- **Type:** Boolean
-- **Default:** `false`
-- **Description:** Enable mock AI providers for testing. Must not be used in production. Useful for development and CI/CD pipelines.
-- **Example:** `export CLEVERAGENTS_MOCK_PROVIDERS=true`
-
-#### `CLEVERAGENTS_TESTING_USE_MOCK_AI`
-- **Type:** Boolean
-- **Default:** `false`
-- **Description:** Force the in-repo mock provider for Behave/Robot test suites to prevent external API calls.
-- **Example:** `export CLEVERAGENTS_TESTING_USE_MOCK_AI=true`
-
-### Logging & Paths
-
-#### `CLEVERAGENTS_LOG_LEVEL`
-- **Type:** String
-- **Default:** `INFO`
-- **Valid values:** `DEBUG`, `INFO`, `WARNING`, `ERROR`, `CRITICAL`
-- **Description:** The logging level for standard output and file logs.
-- **Example:** `export CLEVERAGENTS_LOG_LEVEL=DEBUG`
-
-#### `CLEVERAGENTS_LOG_DIR`
-- **Type:** Path
-- **Default:** `./logs/`
-- **Description:** Directory where log files are written.
-- **Example:** `export CLEVERAGENTS_LOG_DIR=/var/log/cleveragents`
-
-#### `CLEVERAGENTS_DATA_DIR`
-- **Type:** Path
-- **Default:** `~/.cleveragents/`
-- **Description:** Directory for user data, sessions, personas, and other persistent state.
-- **Example:** `export CLEVERAGENTS_DATA_DIR=/data/cleveragents`
-
-#### `CLEVERAGENTS_STORAGE_BASE_PATH`
-- **Type:** Path
-- **Default:** `./data/`
-- **Description:** Base directory for plan storage, checkpoints, and other runtime data.
-- **Example:** `export CLEVERAGENTS_STORAGE_BASE_PATH=/storage/cleveragents`
-
-### Database & Persistence
-
-#### `CLEVERAGENTS_DATABASE_URL`
-- **Type:** String (database URL)
-- **Default:** `sqlite:///$HOME/.cleveragents/cleveragents.db`
-- **Supported formats:**
-  - SQLite: `sqlite:///path/to/database.db`
-  - PostgreSQL: `postgresql://user:password@host:port/database`
-  - MySQL: `mysql+pymysql://user:password@host:port/database`
-  - DuckDB: `duckdb:///path/to/database.duckdb`
-- **Description:** The database URL for storing sessions, plans, and other persistent data.
-- **Example:** `export CLEVERAGENTS_DATABASE_URL=postgresql://user:pass@localhost/cleveragents`
-
-### Cost Controls & Budgets
-
-#### `CLEVERAGENTS_SESSION_MAX_COST_USD`
-- **Type:** Float
-- **Default:** `None` (unlimited)
-- **Valid range:** >= 0.0
-- **Description:** Maximum USD spend per session across all plans. `None` means unlimited.
-- **Example:** `export CLEVERAGENTS_SESSION_MAX_COST_USD=100.00`
-
-#### `CLEVERAGENTS_ORG_MAX_COST_USD`
-- **Type:** Float
-- **Default:** `None` (unlimited)
-- **Valid range:** >= 0.0
-- **Description:** Maximum USD spend per organization across all sessions. `None` means unlimited.
-- **Example:** `export CLEVERAGENTS_ORG_MAX_COST_USD=1000.00`
-
-#### `CLEVERAGENTS_BUDGET_PER_PLAN`
-- **Type:** Float
-- **Default:** `None` (unlimited)
-- **Valid range:** >= 0.0
-- **Description:** Maximum USD spend per plan execution. `None` means unlimited.
-- **Example:** `export CLEVERAGENTS_BUDGET_PER_PLAN=50.00`
-
-#### `CLEVERAGENTS_BUDGET_PER_DAY`
-- **Type:** Float
-- **Default:** `None` (unlimited)
-- **Valid range:** >= 0.0
-- **Description:** Maximum USD spend per calendar day across all plans. `None` means unlimited.
-- **Example:** `export CLEVERAGENTS_BUDGET_PER_DAY=500.00`
-
-### Observability & Logging
-
-#### LangSmith Tracing
-
-#### `CLEVERAGENTS_LANGSMITH_ENABLED`
-- **Type:** Boolean
-- **Default:** `false`
-- **Aliases:** `LANGCHAIN_TRACING_V2`, `LANGSMITH_TRACING_V2`
-- **Description:** Enable LangSmith tracing for LangChain/LangGraph agents.
-- **Example:** `export CLEVERAGENTS_LANGSMITH_ENABLED=true`
-
-#### `CLEVERAGENTS_LANGSMITH_API_KEY`
-- **Type:** String
-- **Default:** `None`
-- **Aliases:** `LANGSMITH_API_KEY`, `LANGCHAIN_API_KEY`
-- **Description:** API key for LangSmith. Required when tracing is enabled.
-- **Example:** `export CLEVERAGENTS_LANGSMITH_API_KEY=ls_...`
-
-#### `CLEVERAGENTS_LANGSMITH_PROJECT`
-- **Type:** String
-- **Default:** `None`
-- **Aliases:** `LANGSMITH_PROJECT`, `LANGCHAIN_PROJECT`
-- **Description:** LangSmith project name. Required when tracing is enabled.
-- **Example:** `export CLEVERAGENTS_LANGSMITH_PROJECT=my-project`
-
-### Cleanup & Retention Policies
-
-#### `CLEVERAGENTS_CLEANUP_SANDBOX_MAX_AGE_HOURS`
-- **Type:** Integer
-- **Default:** `48`
-- **Valid range:** >= 1
-- **Description:** Maximum age (in hours) before stale sandboxes are eligible for cleanup.
-- **Example:** `export CLEVERAGENTS_CLEANUP_SANDBOX_MAX_AGE_HOURS=72`
-
-#### `CLEVERAGENTS_CHECKPOINT_MAX`
-- **Type:** Integer
-- **Default:** `50`
-- **Valid range:** >= 2
-- **Description:** Maximum number of checkpoints to keep per plan. Oldest checkpoints are pruned first, but the first and most recent are always preserved.
-- **Example:** `export CLEVERAGENTS_CHECKPOINT_MAX=100`
-
-### Audit Logging
-
-#### `CLEVERAGENTS_AUDIT_RETENTION_DAYS`
-- **Type:** Integer
-- **Default:** `0` (keep indefinitely)
-- **Valid range:** >= 0
-- **Description:** Days to retain audit log entries before pruning. `0` means keep indefinitely (spec default for compliance).
-- **Example:** `export CLEVERAGENTS_AUDIT_RETENTION_DAYS=365`
-
-#### `CLEVERAGENTS_AUDIT_ASYNC`
-- **Type:** Boolean
-- **Default:** `true`
-- **Description:** When true, audit entries are written asynchronously via a write-behind queue on a background thread. Set to false for synchronous behavior (useful for debugging).
-- **Example:** `export CLEVERAGENTS_AUDIT_ASYNC=false`
-
-### Metrics & Observability
-
-#### `CLEVERAGENTS_METRICS_ENABLED`
-- **Type:** Boolean
-- **Default:** `true`
-- **Description:** Enable structured metric collection and emission.
-- **Example:** `export CLEVERAGENTS_METRICS_ENABLED=false`
-
-#### `CLEVERAGENTS_METRICS_EXPORT_PROMETHEUS`
-- **Type:** Boolean
-- **Default:** `false`
-- **Description:** Enable Prometheus metrics export endpoint.
-- **Example:** `export CLEVERAGENTS_METRICS_EXPORT_PROMETHEUS=true`
-
-### Retry & Circuit Breaker Configuration
-
-#### `CLEVERAGENTS_RETRY_MAX_ATTEMPTS`
-- **Type:** Integer
-- **Default:** `3`
-- **Valid range:** 1-50
-- **Description:** Default maximum retry attempts for service operations.
-- **Example:** `export CLEVERAGENTS_RETRY_MAX_ATTEMPTS=5`
-
-#### `CLEVERAGENTS_RETRY_BASE_DELAY`
-- **Type:** Float
-- **Default:** `1.0`
-- **Valid range:** 0.0-300.0
-- **Description:** Default base delay in seconds between retries.
-- **Example:** `export CLEVERAGENTS_RETRY_BASE_DELAY=0.5`
-
-#### `CLEVERAGENTS_RETRY_MAX_DELAY`
-- **Type:** Float
-- **Default:** `60.0`
-- **Valid range:** 0.0-3600.0
-- **Description:** Default maximum delay in seconds between retries.
-- **Example:** `export CLEVERAGENTS_RETRY_MAX_DELAY=120.0`
-
-#### `CLEVERAGENTS_RETRY_BACKOFF_STRATEGY`
-- **Type:** String
-- **Default:** `exponential`
-- **Valid values:** `exponential`, `linear`, `fixed`, `jitter`, `none`
-- **Description:** Default backoff strategy for retries.
-- **Example:** `export CLEVERAGENTS_RETRY_BACKOFF_STRATEGY=linear`
-
-### Security & Secrets
-
-#### `CLEVERAGENTS_SHOW_SECRETS`
-- **Type:** Boolean
-- **Default:** `false`
-- **Description:** When true, secrets are shown in CLI output and logs. Should never be enabled in production.
-- **Example:** `export CLEVERAGENTS_SHOW_SECRETS=true`
-
-### Automation & Output
-
-#### `CLEVERAGENTS_AUTOMATION_PROFILE`
-- **Type:** String
-- **Default:** `` (empty, manual mode)
-- **Valid values:** `manual`, `auto`, `full-auto` (or custom profile names)
-- **Description:** Default automation profile name controlling plan execution behavior.
-- **Example:** `export CLEVERAGENTS_AUTOMATION_PROFILE=auto`
-
-#### `CLEVERAGENTS_FORMAT`
-- **Type:** String
-- **Default:** `None` (uses default format)
-- **Valid values:** `json`, `markdown`, `text`, `yaml`
-- **Description:** Default output format for CLI commands.
-- **Example:** `export CLEVERAGENTS_FORMAT=json`
-
-## Configuration Best Practices
-
-### Development Environment
-
-For local development, create a `.env` file in your project root:
-
-```bash
-# .env
-CLEVERAGENTS_ENV=development
-CLEVERAGENTS_DEBUG_ENABLED=true
-CLEVERAGENTS_LOG_LEVEL=DEBUG
-OPENAI_API_KEY=sk-proj-...
-CLEVERAGENTS_DATABASE_URL=sqlite:///./dev.db
-```
-
-### Production Environment
-
-For production deployments:
-
-```bash
-# Use strong, randomly generated secrets
-export CLEVERAGENTS_ENV=production
-export CLEVERAGENTS_DEBUG_ENABLED=false
-export CLEVERAGENTS_LOG_LEVEL=WARNING
-export CLEVERAGENTS_SHOW_SECRETS=false
-
-# Use a production database
-export CLEVERAGENTS_DATABASE_URL=postgresql://user:password@db.example.com/cleveragents
-
-# Configure cost controls
-export CLEVERAGENTS_SESSION_MAX_COST_USD=100
-export CLEVERAGENTS_ORG_MAX_COST_USD=1000
-
-# Enable observability
-export CLEVERAGENTS_METRICS_ENABLED=true
-export CLEVERAGENTS_LANGSMITH_ENABLED=true
-export CLEVERAGENTS_LANGSMITH_API_KEY=ls_...
-export CLEVERAGENTS_LANGSMITH_PROJECT=production
-```
-
-### Multi-Provider Setup
-
-To support multiple AI providers with automatic fallback:
-
-```bash
-# Configure multiple providers
-export OPENAI_API_KEY=sk-proj-...
-export ANTHROPIC_API_KEY=sk-ant-...
-export GOOGLE_API_KEY=AIza...
-
-# Set primary provider (optional)
-export CLEVERAGENTS_DEFAULT_PROVIDER=anthropic
-```
-
-## Security Considerations
-
-### Secrets Management
-
-1. **Never commit secrets** — Use environment variables or a secrets manager (Vault, AWS Secrets Manager, etc.)
-2. **Use strong API keys** — Ensure all provider API keys are strong and rotated regularly
-3. **Restrict access** — Limit who can view or modify configuration in your deployment
-4. **Audit logging** — Enable audit logging to track configuration changes and API usage
-5. **Redaction** — CleverAgents automatically redacts secrets in logs and output
-
-### Database Security
-
-- **Use encrypted connections** — Always use TLS/SSL for database connections
-- **Strong credentials** — Use strong, randomly generated database passwords
-- **Network isolation** — Restrict database access to authorized services only
-
-### API Key Security
-
-- **Rotate keys regularly** — Implement automated key rotation (e.g., monthly)
-- **Use scoped keys** — If your provider supports it, use keys with minimal required permissions
-- **Monitor usage** — Track API usage and set up alerts for unusual activity
-
-## Troubleshooting Configuration Issues
-
-### No AI Provider Configured
-
-**Error:** `ValueError: No AI providers configured and mock_providers is disabled.`
-
-**Solution:**
-1. Set at least one provider API key: `export OPENAI_API_KEY=sk-proj-...`
-2. Or enable mock providers for testing: `export CLEVERAGENTS_MOCK_PROVIDERS=true`
-3. Verify the key is set: `echo $OPENAI_API_KEY`
-
-### LangSmith Configuration Incomplete
-
-**Error:** `LangSmith API key is required` or `LangSmith project name is required`
-
-**Solution:**
-1. Set both required variables:
-   ```bash
-   export CLEVERAGENTS_LANGSMITH_API_KEY=ls_...
-   export CLEVERAGENTS_LANGSMITH_PROJECT=my-project
-   ```
-2. Or disable LangSmith: `export CLEVERAGENTS_LANGSMITH_ENABLED=false`
-
-### Database Connection Failed
-
-**Error:** `sqlalchemy.exc.OperationalError: (psycopg2.OperationalError) could not connect to server`
-
-**Solution:**
-1. Verify the database URL: `echo $CLEVERAGENTS_DATABASE_URL`
-2. Check database credentials and host
-3. Ensure the database is running and accessible
-
-### Configuration Not Taking Effect
-
-**Problem:** Environment variable changes not reflected
-
-**Solution:**
-1. Verify the variable is set: `env | grep CLEVERAGENTS`
-2. Restart the application (singleton settings are cached)
-3. Check for typos in variable names (case-sensitive)
-4. Ensure the variable is exported: `export CLEVERAGENTS_VAR=value`
-
-## Related Documentation
-
-- [Architecture Guide](../architecture.md) — System design and component overview
-- [API Reference - Configuration](../api/config.md) — Detailed API documentation
-- [Development Guide](../development/quality-automation.md) — Development setup and workflows
-- [Specification](../specification.md) — Authoritative design documentation
-- [FAQ](../faq.md) — Frequently asked questions

docs/guides/context_management_deep_dive.md

@@ -0,0 +1,998 @@
+# Context Management Deep Dive
+
+This comprehensive guide explores the Adaptive Context Management System (ACMS) architecture, strategies, and best practices for managing context in cleveragents-core.
+
+## Table of Contents
+
+1. [Context Management Architecture Overview](#context-management-architecture-overview)
+2. [ACMS Pipeline Components](#acms-pipeline-components)
+3. [Context Strategy Patterns](#context-strategy-patterns)
+4. [Context Inheritance and Scope Resolution](#context-inheritance-and-scope-resolution)
+5. [Budget Allocation and Utilization Metrics](#budget-allocation-and-utilization-metrics)
+6. [Implementation Guide for Custom Strategies](#implementation-guide-for-custom-strategies)
+7. [Best Practices and Performance Considerations](#best-practices-and-performance-considerations)
+8. [Integration with Plan Execution](#integration-with-plan-execution)
+9. [Real-world Examples](#real-world-examples)
+
+## Context Management Architecture Overview
+
+### What is Context Management?
+
+Context management is the process of selecting, prioritizing, and packaging relevant information from a project's knowledge base to provide to language models (LMs) during plan execution. The ACMS ensures that:
+
+- **Relevance**: Only the most pertinent information is included
+- **Budget Compliance**: Token limits are strictly respected
+- **Efficiency**: Context assembly is fast and deterministic
+- **Traceability**: All context is provenance-tracked back to its source
+
+### The Three-Tier Context Model
+
+ACMS organizes context into three priority tiers:
+
+| Tier | Purpose | Inclusion Strategy |
+|------|---------|-------------------|
+| **hot** | Critical context that must be included | Always included first; highest priority |
+| **warm** | Standard context included when budget allows | Included after hot tier; default tier |
+| **cold** | Low-priority context included only if space remains | Included last; only if budget permits |
+
+> **Note:** In v1, tiers are sort-priority labels for ranking during assembly. Full storage-tier semantics with retention policies and promotion/demotion will be implemented in future milestones.
+
+### Core Concepts
+
+#### Context Fragments
+
+A **ContextFragment** is the atomic unit of context. Each fragment represents a single piece of information:
+
+```python
+ContextFragment(
+    uko_node="project://myapp/src/main.py",      # UKO URI of source
+    content="def main(): ...",                     # Rendered text content
+    detail_depth=3,                                # Depth level (0-9)
+    token_count=100,                               # Actual token count
+    relevance_score=0.9,                           # Relevance (0.0-1.0)
+    tier="hot",                                    # Priority tier
+    provenance=FragmentProvenance(                 # Source tracking
+        resource_uri="project://myapp/src/main.py",
+        location="line 1-10",
+        resource_type="source_file"
+    )
+)
+```
+
+**Key Properties:**
+- `fragment_id`: Unique ULID identifier (auto-generated)
+- `content`: Max 1,000,000 characters
+- `detail_depth`: Resolved depth from 0 (MODULE_LISTING) to 9 (FULL_SOURCE)
+- `token_count`: Must be >= 0; used for budget calculations
+- `relevance_score`: 0.0 (irrelevant) to 1.0 (highly relevant)
+- `metadata`: Arbitrary key-value pairs (max 64 entries)
+- `created_at`: UTC timestamp for recency-based sorting
+
+#### Context Budget
+
+The **ContextBudget** defines token constraints:
+
+```python
+budget = ContextBudget(
+    max_tokens=4096,           # Total available tokens
+    reserved_tokens=512        # Tokens reserved for system prompt
+)
+# Available tokens = max_tokens - reserved_tokens = 3584
+```
+
+**Budget Rules:**
+- `available_tokens = max_tokens - reserved_tokens`
+- `reserved_tokens` must be < `max_tokens`
+- Pipeline never exceeds `available_tokens`
+- Fragments exceeding budget are skipped
+
+#### Context Payload
+
+The assembled **ContextPayload** is the final output:
+
+```python
+payload = ContextPayload(
+    payload_id="01ARZ3NDEKTSV4RRFFQ69G5FAV",     # ULID
+    plan_id="plan-1",                             # Associated plan
+    fragments=(fragment1, fragment2, ...),        # Selected fragments
+    total_tokens=2048,                            # Sum of fragment tokens
+    budget=budget,                                # Budget used
+    budget_used=0.57,                             # Fraction consumed
+    strategies_used=("simple-keyword",),          # Contributing strategies
+    context_hash="abc123...",                     # SHA-256 hash
+    preamble="Context includes 2 files...",       # Optional summary
+    provenance_map={...},                         # Fragment ID -> provenance
+    assembled_at=datetime.now(timezone.utc)       # Assembly timestamp
+)
+```
+
+**Key Properties:**
+- `is_within_budget`: True if total_tokens <= available_tokens
+- `remaining_tokens`: available_tokens - total_tokens
+
+### Context Request Protocol (CRP)
+
+The **ContextRequest** specifies what context is needed:
+
+```python
+request = ContextRequest(
+    focus_nodes=["project://myapp/src/main.py"],  # Primary focus
+    breadth=2,                                     # Graph traversal breadth
+    depth=3,                                       # Graph traversal depth
+    keywords=["async", "io"],                      # Search keywords
+    resource_types=["source_file", "decision"],    # Filter by type
+    detail_level=3,                                # Desired detail depth
+    temporal_window="7d",                          # Historical window
+    strategy_hints=["semantic-embedding"],         # Preferred strategies
+)
+```
+
+## ACMS Pipeline Components
+
+The ACMS pipeline implements a 10-component architecture across three phases:
+
+### Phase 1: Strategy Orchestration
+
+#### 1. StrategySelector
+
+**Purpose**: Determine which strategies can handle the request.
+
+**Process**:
+1. Poll all registered strategies with `can_handle(request)`
+2. Each strategy returns confidence (0.0-1.0)
+3. Select strategies above confidence threshold
+4. Rank by confidence score
+
+#### 2. BudgetAllocator
+
+**Purpose**: Distribute token budget among selected strategies.
+
+**Allocation Strategies**:
+- **Proportional**: Allocate based on confidence scores
+- **Equal**: Divide equally among strategies
+- **Priority**: Allocate to highest-confidence first
+
+#### 3. StrategyExecutor
+
+**Purpose**: Execute strategies in parallel with timeout and circuit breaker.
+
+**Features**:
+- Parallel execution with configurable workers
+- Per-strategy timeout (default 30s)
+- Circuit breaker after N failures
+- Graceful degradation on timeout
+
+### Phase 2: Fragment Fusion
+
+#### 4. FragmentDeduplicator
+
+**Purpose**: Remove duplicate fragments based on UKO node and content hash.
+
+**Deduplication Rules**:
+- Same `uko_node` + same content hash = duplicate
+- Keep fragment with highest relevance score
+- Preserve provenance from all duplicates
+
+#### 5. DetailDepthResolver
+
+**Purpose**: Resolve appropriate detail depth for each fragment.
+
+**Resolution Logic**:
+- Start with requested detail level
+- Adjust based on token budget
+- Respect fragment's max depth
+- Prefer higher depth for hot tier
+
+#### 6. FragmentScorer
+
+**Purpose**: Score fragments based on multiple factors.
+
+**Scoring Factors**:
+- Relevance score (from strategy)
+- Tier priority (hot > warm > cold)
+- Recency (created_at)
+- Plan context alignment
+- Resource type affinity
+
+#### 7. BudgetPacker
+
+**Purpose**: Select fragments to fit within token budget.
+
+**Packing Algorithm**:
+1. Sort fragments by score (descending)
+2. Greedily add fragments until budget exhausted
+3. Respect tier ordering (hot → warm → cold)
+4. Early termination when budget full
+
+#### 8. FragmentOrderer
+
+**Purpose**: Determine final fragment order in payload.
+
+**Ordering Strategies**:
+- **Relevance**: By relevance_score (descending)
+- **Recency**: By created_at (descending)
+- **Tier-Relevance**: By tier, then relevance
+- **Custom**: User-defined ordering
+
+### Phase 3: Context Finalization
+
+#### 9. PreambleGenerator
+
+**Purpose**: Generate optional summary of assembled context.
+
+**Preamble Contents**:
+- Number of fragments and sources
+- Token usage and budget utilization
+- Strategies used
+- Key topics covered
+- Temporal range
+
+#### 10. SkeletonCompressor
+
+**Purpose**: Compress context for child plans while preserving critical information.
+
+**Compression Strategy**:
+- Identify critical fragments for child focus
+- Remove redundant details
+- Preserve decision context
+- Maintain provenance
+
+## Context Strategy Patterns
+
+### Strategy Architecture
+
+All strategies implement the `ContextStrategy` protocol:
+
+```python
+class ContextStrategy(Protocol):
+    @property
+    def name(self) -> str:
+        """Unique strategy identifier"""
+        ...
+    
+    @property
+    def capabilities(self) -> StrategyCapabilities:
+        """Declare backend requirements and quality"""
+        ...
+    
+    def can_handle(
+        self,
+        request: ContextRequest,
+        backends: BackendSet
+    ) -> float:
+        """Return confidence (0.0-1.0) for handling request"""
+        ...
+    
+    def assemble(
+        self,
+        request: ContextRequest,
+        backends: BackendSet,
+        budget: ContextBudget,
+        plan_context: PlanContext
+    ) -> list[ContextFragment]:
+        """Execute strategy and return fragments"""
+        ...
+    
+    def explain(self) -> str:
+        """Human-readable description"""
+        ...
+```
+
+### Built-in Strategies
+
+#### 1. Simple Keyword Strategy
+
+**Quality Score**: 0.3 (baseline)
+
+**Backend Requirements**: TextBackend
+
+**How It Works**:
+1. Extract keywords from request
+2. Search TextBackend using keywords
+3. Rank results by match frequency
+4. Pack greedily into budget
+
+**Best For**:
+- Quick searches with clear keywords
+- Projects with good text indexing
+- Fallback when other backends unavailable
+
+#### 2. Semantic Embedding Strategy
+
+**Quality Score**: 0.6 (good)
+
+**Backend Requirements**: VectorBackend
+
+**How It Works**:
+1. Embed request using character-frequency model (v1)
+2. Search VectorBackend for similar embeddings
+3. Rank by cosine similarity
+4. Pack budget-aware results
+
+**Best For**:
+- Semantic similarity searches
+- Conceptual relationships
+- Cross-domain context discovery
+
+#### 3. Breadth-Depth Navigator Strategy
+
+**Quality Score**: 0.85 (excellent)
+
+**Backend Requirements**: GraphBackend
+
+**How It Works**:
+1. Start from focus nodes in request
+2. Traverse graph outward by breadth hops
+3. Traverse downward by depth hops
+4. Collect all reachable nodes
+5. Rank by distance and relevance
+6. Pack into budget
+
+**Best For**:
+- Exploring related code and decisions
+- Understanding dependencies
+- Comprehensive context assembly
+
+#### 4. ARCE Strategy (Advanced Retrieval with Context Enrichment)
+
+**Quality Score**: 0.95 (highest)
+
+**Backend Requirements**: TextBackend, VectorBackend, GraphBackend
+
+**How It Works**:
+1. **Phase 1 (40% budget)**: Text search using keywords
+2. **Phase 2 (40% budget)**: Vector similarity search
+3. **Phase 3 (20% budget)**: Graph traversal from results
+4. Merge and deduplicate results
+5. Rank by combined score
+6. Pack into budget
+
+**Best For**:
+- Comprehensive context assembly
+- Multi-modal queries
+- Production use cases
+
+#### 5. Temporal Archaeology Strategy
+
+**Quality Score**: 0.5 (moderate)
+
+**Backend Requirements**: GraphBackend, TemporalBackend (cold tier)
+
+**How It Works**:
+1. Query TemporalBackend for historical nodes
+2. Filter by temporal window
+3. Traverse graph from historical nodes
+4. Collect related current nodes
+5. Rank by recency and relevance
+6. Pack into budget
+
+**Best For**:
+- Understanding historical context
+- Decision archaeology
+- Long-term project evolution
+
+#### 6. Plan-Decision Context Strategy
+
+**Quality Score**: 0.7 (good)
+
+**Backend Requirements**: TemporalBackend (warm/cold), GraphBackend
+
+**How It Works**:
+1. Walk parent and ancestor plan hierarchy
+2. Retrieve decision records from warm/cold tiers
+3. Traverse graph from decisions
+4. Collect related context
+5. Rank by plan proximity and relevance
+6. Pack into budget
+
+**Best For**:
+- Understanding plan context
+- Decision inheritance
+- Multi-level plan coordination
+
+### Strategy Comparison Matrix
+
+| Strategy | Quality | Speed | Coverage | Best For |
+|----------|---------|-------|----------|----------|
+| Simple Keyword | 0.3 | ⚡⚡⚡ | Narrow | Quick searches |
+| Semantic Embedding | 0.6 | ⚡⚡ | Broad | Semantic similarity |
+| Breadth-Depth Navigator | 0.85 | ⚡ | Very Broad | Comprehensive context |
+| ARCE | 0.95 | 🐢 | Comprehensive | Production use |
+| Temporal Archaeology | 0.5 | 🐢 | Historical | Historical context |
+| Plan-Decision Context | 0.7 | ⚡ | Hierarchical | Plan coordination |
+
+## Context Inheritance and Scope Resolution
+
+### Scope Hierarchy
+
+Context is organized in a hierarchy of scopes:
+
+```
+Global Scope
+  ├── Project Scope
+  │   ├── Module Scope
+  │   │   ├── Class Scope
+  │   │   │   └── Method Scope
+  │   │   └── Function Scope
+  │   └── Decision Scope
+  └── Temporal Scope
+      ├── Current
+      ├── Recent (7d)
+      └── Historical (30d+)
+```
+
+### Scope Resolution Rules
+
+When assembling context for a request:
+
+1. **Start at Request Focus**: Begin with explicitly requested nodes
+2. **Expand Upward**: Include parent scopes (module → project → global)
+3. **Expand Downward**: Include child scopes (class → methods)
+4. **Expand Sideways**: Include sibling scopes (related classes)
+5. **Apply Filters**: Remove out-of-scope items based on resource types
+6. **Respect Budget**: Stop expansion when budget exhausted
+
+### Context Inheritance
+
+Child plans inherit context from parent plans:
+
+```python
+# Parent plan context
+parent_payload = ContextPayload(
+    fragments=[
+        ContextFragment(uko_node="project://myapp/src/main.py", ...),
+        ContextFragment(uko_node="project://myapp/src/utils.py", ...),
+        ContextFragment(uko_node="project://myapp/decisions/001", ...),
+    ]
+)
+
+# Child plan inherits and specializes
+child_request = ContextRequest(
+    focus_nodes=["project://myapp/src/main.py"],
+    parent_context=parent_payload,
+    inherit_strategy="skeleton"  # Compress parent context
+)
+
+# Result: Child gets focused context + compressed parent context
+child_payload = pipeline.assemble(child_request)
+```
+
+**Inheritance Strategies**:
+- **Full**: Include all parent context (if budget allows)
+- **Skeleton**: Compress parent context to essentials
+- **None**: Start fresh (no parent context)
+
+## Budget Allocation and Utilization Metrics
+
+### Budget Planning
+
+Effective budget allocation requires understanding token costs:
+
+```python
+# Typical token costs
+costs = {
+    "system_prompt": 512,           # Reserved
+    "user_request": 200,            # Typical request
+    "model_response": 1024,         # Expected response
+    "context": 2048,                # Available for context
+}
+
+total_budget = 4096
+reserved = 512
+available = total_budget - reserved  # 3584 tokens
+```
+
+### Budget Allocation Strategies
+
+#### Proportional Allocation
+
+Allocate based on strategy confidence scores:
+
+```python
+candidates = [
+    (semantic_embedding, 0.95),
+    (breadth_depth, 0.85),
+    (simple_keyword, 0.3)
+]
+
+total_confidence = 0.95 + 0.85 + 0.3  # 2.1
+allocations = {
+    semantic_embedding: int(3584 * 0.95 / 2.1),   # 1621
+    breadth_depth: int(3584 * 0.85 / 2.1),        # 1450
+    simple_keyword: int(3584 * 0.3 / 2.1),        # 513
+}
+```
+
+#### Priority Allocation
+
+Allocate to highest-confidence strategies first:
+
+```python
+allocations = {}
+remaining = 3584
+
+for strategy, confidence in sorted(candidates, key=lambda x: x[1], reverse=True):
+    # Allocate 50% of remaining budget
+    allocation = remaining // 2
+    allocations[strategy] = allocation
+    remaining -= allocation
+```
+
+#### Tier-Based Allocation
+
+Allocate based on tier requirements:
+
+```python
+allocations = {
+    "hot_tier": 1000,      # 28% - Critical context
+    "warm_tier": 2000,     # 56% - Standard context
+    "cold_tier": 584,      # 16% - Low-priority context
+}
+```
+
+### Utilization Metrics
+
+Track context assembly efficiency:
+
+```python
+metrics = {
+    "budget_used": payload.budget_used,              # 0.57 (57%)
+    "remaining_tokens": payload.remaining_tokens,    # 1536
+    "fragments_selected": len(payload.fragments),    # 12
+    "fragments_total": 45,                           # Total available
+    "coverage": 12 / 45,                             # 26.7%
+    "strategies_used": payload.strategies_used,      # ("semantic-embedding",)
+    "assembly_time_ms": 234,                         # Execution time
+    "avg_fragment_tokens": 2048 / 12,                # 170.7
+    "avg_relevance": 0.82,                           # Mean relevance score
+}
+```
+
+### Budget Optimization
+
+Techniques to improve budget utilization:
+
+1. **Increase Reserved Tokens**: Allocate more for system prompt
+2. **Reduce Detail Depth**: Lower detail_depth for less critical fragments
+3. **Increase Relevance Threshold**: Filter low-relevance fragments
+4. **Use Skeleton Compression**: Compress parent context for child plans
+5. **Tier-Based Filtering**: Exclude cold tier if budget tight
+
+## Implementation Guide for Custom Strategies
+
+### Creating a Custom Strategy
+
+#### Step 1: Define Strategy Class
+
+```python
+from cleveragents.application.services.acms_service import (
+    ContextStrategy,
+    StrategyCapabilities,
+)
+from cleveragents.domain.models.core.context_fragment import (
+    ContextBudget,
+    ContextFragment,
+)
+from typing import Sequence, Any
+
+
+class MyCustomStrategy:
+    """Custom context strategy for specialized retrieval."""
+    
+    @property
+    def name(self) -> str:
+        """Unique strategy identifier."""
+        return "my-custom-strategy"
+    
+    @property
+    def capabilities(self) -> StrategyCapabilities:
+        """Declare backend requirements and quality."""
+        return StrategyCapabilities(
+            uses_text=True,
+            uses_vector=False,
+            uses_graph=True,
+            uses_temporal=False,
+            resource_types=("source_file", "decision"),
+            quality_score=0.75,
+            supports_depth_breadth=True,
+            supports_plan_hierarchy=False,
+            supports_temporal=False,
+        )
+    
+    def can_handle(self, request: dict[str, Any]) -> float:
+        """Return confidence (0.0-1.0) for handling request."""
+        # Check if request has required fields
+        if "focus_nodes" not in request:
+            return 0.0
+        
+        # Check for specific keywords
+        keywords = request.get("keywords", [])
+        if any(kw in ["async", "io", "performance"] for kw in keywords):
+            return 0.9  # High confidence for performance-related queries
+        
+        return 0.6  # Moderate confidence for general queries
+    
+    def assemble(
+        self,
+        fragments: Sequence[ContextFragment],
+        budget: ContextBudget,
+    ) -> Sequence[ContextFragment]:
+        """Execute strategy and return fragments."""
+        # Custom ranking logic
+        ranked = sorted(
+            fragments,
+            key=lambda f: (
+                f.tier == "hot",           # Hot tier first
+                f.relevance_score,         # Then by relevance
+                f.created_at,              # Then by recency
+            ),
+            reverse=True
+        )
+        
+        # Pack into budget
+        selected = []
+        total_tokens = 0
+        
+        for fragment in ranked:
+            if total_tokens + fragment.token_count <= budget.available_tokens:
+                selected.append(fragment)
+                total_tokens += fragment.token_count
+            else:
+                break  # Budget exhausted
+        
+        return selected
+    
+    def explain(self) -> str:
+        """Human-readable description."""
+        return (
+            "Custom strategy that prioritizes hot-tier fragments "
+            "and performance-related keywords."
+        )
+```
+
+#### Step 2: Register Strategy
+
+```python
+from cleveragents.application.services.acms_service import ACMSPipeline
+
+# Create pipeline
+pipeline = ACMSPipeline()
+
+# Register custom strategy
+strategy = MyCustomStrategy()
+pipeline.register_strategy(strategy.name, strategy)
+
+# Use in assembly
+payload = pipeline.assemble(
+    plan_id="plan-1",
+    fragments=fragments,
+    budget=budget,
+    strategy="my-custom-strategy"
+)
+```
+
+#### Step 3: Configure in TOML
+
+```toml
+[context.strategies]
+enabled = ["simple-keyword", "my-custom-strategy"]
+
+[context.strategies.custom]
+"my-custom-strategy" = "my_package.strategies:MyCustomStrategy"
+
+[context.strategies."my-custom-strategy"]
+enabled = true
+timeout_seconds = 30
+max_fragments = 100
+```
+
+## Best Practices and Performance Considerations
+
+### Performance Optimization
+
+#### 1. Strategy Selection
+
+**Best Practice**: Use strategy confidence scores to avoid unnecessary execution.
+
+#### 2. Budget Allocation
+
+**Best Practice**: Allocate more budget to higher-confidence strategies.
+
+#### 3. Timeout Management
+
+**Best Practice**: Set appropriate timeouts based on strategy complexity.
+
+#### 4. Fragment Deduplication
+
+**Best Practice**: Deduplicate early to reduce downstream processing.
+
+#### 5. Detail Depth Resolution
+
+**Best Practice**: Reduce detail depth for less critical fragments.
+
+### Memory Efficiency
+
+#### 1. Fragment Streaming
+
+**Best Practice**: Stream fragments instead of loading all at once.
+
+#### 2. Lazy Evaluation
+
+**Best Practice**: Defer expensive operations until needed.
+
+### Reliability and Robustness
+
+#### 1. Circuit Breaker Pattern
+
+**Best Practice**: Implement circuit breaker for failing strategies.
+
+#### 2. Graceful Degradation
+
+**Best Practice**: Fall back to simpler strategies when complex ones fail.
+
+#### 3. Timeout Handling
+
+**Best Practice**: Set reasonable timeouts with fallback.
+
+### Monitoring and Observability
+
+#### 1. Metrics Collection
+
+**Best Practice**: Collect comprehensive metrics for monitoring.
+
+#### 2. Logging
+
+**Best Practice**: Log key decisions and failures.
+
+## Integration with Plan Execution
+
+### Context Assembly in Plan Execution
+
+Context is assembled at key points in plan execution:
+
+```
+Plan Execution Flow
+├── Plan Created
+│   └── Initial Context Assembly
+│       ├── StrategySelector: Choose strategies
+│       ├── BudgetAllocator: Allocate tokens
+│       └── StrategyExecutor: Execute strategies
+├── Plan Step Execution
+│   └── Context Refresh (optional)
+│       └── Incremental context updates
+├── Subplan Creation
+│   └── Context Inheritance
+│       ├── Inherit parent context
+│       └── Specialize for subplan focus
+└── Plan Completion
+    └── Context Archive
+        └── Store for historical reference
+```
+
+### Context Request from Plan
+
+Plans generate context requests based on their focus.
+
+### Context Injection into Plan
+
+Assembled context is injected into plan execution.
+
+### Subplan Context Inheritance
+
+Child plans inherit and specialize context.
+
+## Real-world Examples
+
+### Example 1: Implementing a New Feature
+
+**Scenario**: Implement async I/O support in a Python project.
+
+**Context Assembly**:
+
+```python
+# 1. Create context request
+request = ContextRequest(
+    focus_nodes=[
+        "project://myapp/src/main.py",
+        "project://myapp/src/io.py"
+    ],
+    keywords=["async", "io", "performance"],
+    breadth=2,
+    depth=3,
+    resource_types=["source_file", "decision"],
+    detail_level=4
+)
+
+# 2. Assemble context
+payload = pipeline.assemble(
+    plan_id="plan-async-io",
+    fragments=fetch_fragments(backends, request),
+    budget=ContextBudget(max_tokens=4096, reserved_tokens=512),
+    strategy="arce"  # Use comprehensive strategy
+)
+
+# 3. Inspect assembled context
+print(f"Fragments: {len(payload.fragments)}")
+print(f"Tokens: {payload.total_tokens} / {payload.budget.available_tokens}")
+print(f"Budget used: {payload.budget_used:.1%}")
+print(f"Strategies: {payload.strategies_used}")
+
+# 4. Use context in plan execution
+plan = Plan(
+    title="Implement async I/O",
+    context_payload=payload,
+    context_hash=payload.context_hash
+)
+```
+
+**Result**: LM receives focused context on async I/O patterns, related decisions, and performance considerations.
+
+### Example 2: Debugging a Performance Issue
+
+**Scenario**: Investigate slow database queries.
+
+**Context Assembly**:
+
+```python
+# 1. Create context request focused on database code
+request = ContextRequest(
+    focus_nodes=["project://myapp/src/database.py"],
+    keywords=["query", "performance", "slow", "index"],
+    breadth=3,  # Broader search for related code
+    depth=2,    # Less deep (focus on structure)
+    resource_types=["source_file", "decision", "test"],
+    detail_level=2,  # Less detail to fit more context
+    temporal_window="30d"  # Include recent changes
+)
+
+# 2. Assemble with temporal strategy
+payload = pipeline.assemble(
+    plan_id="plan-debug-perf",
+    fragments=fetch_fragments(backends, request),
+    budget=ContextBudget(max_tokens=4096, reserved_tokens=512),
+    strategy="temporal-archaeology"  # Include historical context
+)
+
+# 3. Analyze context
+hot_fragments = [f for f in payload.fragments if f.tier == "hot"]
+recent_fragments = sorted(
+    payload.fragments,
+    key=lambda f: f.created_at,
+    reverse=True
+)[:5]
+
+print(f"Hot tier fragments: {len(hot_fragments)}")
+print(f"Recent changes: {[f.uko_node for f in recent_fragments]}")
+```
+
+**Result**: LM receives context on database code, recent changes, and historical decisions about performance.
+
+### Example 3: Code Review with Context
+
+**Scenario**: Review a pull request with full context.
+
+**Context Assembly**:
+
+```python
+# 1. Get changed files from PR
+changed_files = [
+    "project://myapp/src/auth.py",
+    "project://myapp/src/security.py",
+    "project://myapp/tests/test_auth.py"
+]
+
+# 2. Create context request
+request = ContextRequest(
+    focus_nodes=changed_files,
+    breadth=1,  # Narrow search (focus on changed files)
+    depth=4,    # Deeper (understand full context)
+    resource_types=["source_file", "decision", "test"],
+    detail_level=5,  # Higher detail for review
+    strategy_hints=["breadth-depth-navigator"]  # Use graph traversal
+)
+
+# 3. Assemble context
+payload = pipeline.assemble(
+    plan_id="plan-review-pr",
+    fragments=fetch_fragments(backends, request),
+    budget=ContextBudget(max_tokens=8192, reserved_tokens=512),
+    strategy="breadth-depth-navigator"
+)
+
+# 4. Generate review context
+review_context = {
+    "changed_files": changed_files,
+    "context_fragments": len(payload.fragments),
+    "context_tokens": payload.total_tokens,
+    "related_decisions": [
+        f for f in payload.fragments
+        if "decision" in f.metadata.get("resource_type", "")
+    ],
+    "related_tests": [
+        f for f in payload.fragments
+        if "test" in f.uko_node
+    ]
+}
+```
+
+**Result**: LM receives comprehensive context on changed files, related code, decisions, and tests for thorough review.
+
+### Example 4: Multi-level Plan Hierarchy
+
+**Scenario**: Execute a complex plan with multiple subplans.
+
+**Context Assembly**:
+
+```python
+# Parent plan context
+parent_request = ContextRequest(
+    focus_nodes=["project://myapp"],
+    breadth=2,
+    depth=2,
+    detail_level=2
+)
+
+parent_payload = pipeline.assemble(
+    plan_id="plan-refactor",
+    fragments=fetch_fragments(backends, parent_request),
+    budget=ContextBudget(max_tokens=4096, reserved_tokens=512),
+    strategy="arce"
+)
+
+# Subplan 1: Refactor auth module
+subplan1_request = ContextRequest(
+    focus_nodes=["project://myapp/src/auth.py"],
+    parent_context=parent_payload,
+    inherit_strategy="skeleton",  # Compress parent
+    breadth=2,
+    depth=3,
+    detail_level=4
+)
+
+subplan1_payload = pipeline.assemble(
+    plan_id="plan-refactor-auth",
+    fragments=fetch_fragments(backends, subplan1_request),
+    budget=ContextBudget(max_tokens=3584, reserved_tokens=512),
+    strategy="breadth-depth-navigator"
+)
+
+# Subplan 2: Refactor database module
+subplan2_request = ContextRequest(
+    focus_nodes=["project://myapp/src/database.py"],
+    parent_context=parent_payload,
+    inherit_strategy="skeleton",
+    breadth=2,
+    depth=3,
+    detail_level=4
+)
+
+subplan2_payload = pipeline.assemble(
+    plan_id="plan-refactor-db",
+    fragments=fetch_fragments(backends, subplan2_request),
+    budget=ContextBudget(max_tokens=3584, reserved_tokens=512),
+    strategy="breadth-depth-navigator"
+)
+
+# Execute hierarchy
+execute_plan(parent_plan, parent_payload)
+execute_plan(subplan1, subplan1_payload)
+execute_plan(subplan2, subplan2_payload)
+```
+
+**Result**: Parent plan gets broad context; subplans get specialized context with parent context compressed for efficiency.
+
+## Conclusion
+
+The ACMS provides a powerful, flexible system for managing context in cleveragents-core. By understanding the architecture, strategies, and best practices, you can:
+
+- **Assemble relevant context** efficiently and deterministically
+- **Respect token budgets** while maximizing information density
+- **Implement custom strategies** for specialized use cases
+- **Optimize performance** through careful budget allocation
+- **Integrate seamlessly** with plan execution
+
+For more information, see:
+- [Context Strategies Reference](../reference/context_strategies_reference.md)
+- [ACMS Pipeline Reference](../reference/acms_pipeline_reference.md)
+- [ACMS v1 Context Assembly Pipeline](../reference/acms.md)
+- [Context Strategy Registry](../reference/context_strategies.md)

docs/reference/acms_pipeline_reference.md

@@ -0,0 +1,760 @@
+# ACMS Pipeline Reference
+
+Complete reference for the ACMS Pipeline components, their interactions, data flow, configuration, and monitoring.
+
+## Pipeline Architecture
+
+The ACMS pipeline implements a 10-component architecture across three phases:
+
+```
+Phase 1: Strategy Orchestration
+├── StrategySelector
+├── BudgetAllocator
+└── StrategyExecutor
+
+Phase 2: Fragment Fusion
+├── FragmentDeduplicator
+├── DetailDepthResolver
+├── FragmentScorer
+├── BudgetPacker
+└── FragmentOrderer
+
+Phase 3: Context Finalization
+├── PreambleGenerator
+└── SkeletonCompressor
+```
+
+## Component Reference
+
+### Phase 1: Strategy Orchestration
+
+#### StrategySelector
+
+**Purpose**: Determine which strategies can handle the request.
+
+**Input**:
+- `strategies`: List of registered strategies
+- `request`: ContextRequest with focus nodes, keywords, etc.
+- `backends`: Available backends (TextBackend, VectorBackend, etc.)
+
+**Output**:
+- `candidates`: List of (strategy, confidence) tuples, sorted by confidence descending
+
+**Process**:
+1. Poll each strategy with `can_handle(request, backends)`
+2. Collect confidence scores (0.0-1.0)
+3. Filter strategies with confidence >= threshold (default 0.0)
+4. Sort by confidence descending
+5. Return top candidates
+
+**Configuration**:
+```toml
+[context.strategies]
+selector_threshold = 0.0
+max_candidates = 10
+```
+
+**Example**:
+```python
+selector = StrategySelector()
+candidates = selector.select(
+    strategies=[simple_keyword, semantic_embedding, breadth_depth],
+    request=context_request,
+    backends=backend_set
+)
+# Returns: [(semantic_embedding, 0.95), (breadth_depth, 0.85), ...]
+```
+
+#### BudgetAllocator
+
+**Purpose**: Distribute token budget among selected strategies.
+
+**Input**:
+- `candidates`: List of (strategy, confidence) tuples
+- `total_budget`: Available tokens for context
+- `request`: ContextRequest with budget hints
+
+**Output**:
+- `allocations`: Dict mapping strategy -> allocated tokens
+
+**Allocation Strategies**:
+
+1. **Proportional** (default):
+   ```
+   allocation[s] = total_budget * confidence[s] / sum(confidences)
+   ```
+
+2. **Equal**:
+   ```
+   allocation[s] = total_budget / len(candidates)
+   ```
+
+3. **Priority**:
+   ```
+   Sort by confidence, allocate 50% of remaining to each
+   ```
+
+**Configuration**:
+```toml
+[context.strategies]
+allocator_strategy = "proportional"
+allocator_min_budget = 100
+```
+
+**Example**:
+```python
+allocator = BudgetAllocator()
+allocations = allocator.allocate(
+    candidates=[(semantic_embedding, 0.95), (breadth_depth, 0.85)],
+    total_budget=3584,
+    request=context_request
+)
+# Returns: {semantic_embedding: 1900, breadth_depth: 1684}
+```
+
+#### StrategyExecutor
+
+**Purpose**: Execute strategies in parallel with timeout and circuit breaker.
+
+**Input**:
+- `allocations`: Dict mapping strategy -> allocated tokens
+- `request`: ContextRequest
+- `backends`: Available backends
+- `plan_context`: Current plan context
+
+**Output**:
+- `results`: List of StrategyResult objects
+
+**Features**:
+- Parallel execution with configurable workers
+- Per-strategy timeout (default 30s)
+- Circuit breaker after N failures
+- Graceful degradation on timeout
+
+**Configuration**:
+```toml
+[context.strategies]
+executor_max_workers = 4
+executor_timeout_seconds = 30
+executor_circuit_breaker_threshold = 3
+executor_circuit_breaker_reset_timeout = 60
+```
+
+**Example**:
+```python
+executor = StrategyExecutor(max_workers=4, timeout_seconds=30)
+results = executor.execute(
+    allocations={semantic_embedding: 1900, breadth_depth: 1684},
+    request=context_request,
+    backends=backend_set,
+    plan_context=plan_context
+)
+# Returns: [StrategyResult(...), StrategyResult(...)]
+```
+
+**StrategyResult**:
+```python
+@dataclass
+class StrategyResult:
+    strategy_name: str
+    fragments: tuple[ContextFragment, ...]
+    total_fragments: int
+    tokens_used: int
+    execution_time_ms: float
+    errors: tuple[str, ...]
+    stats: dict[str, Any]
+```
+
+### Phase 2: Fragment Fusion
+
+#### FragmentDeduplicator
+
+**Purpose**: Remove duplicate fragments based on UKO node and content hash.
+
+**Input**:
+- `fragments`: List of ContextFragment objects
+- `strategy`: Deduplication strategy ("highest_relevance", "first_seen", "most_recent")
+
+**Output**:
+- `deduplicated`: List of unique fragments
+
+**Deduplication Rules**:
+- Same `uko_node` + same content hash = duplicate
+- Keep fragment with highest relevance score (default)
+- Merge provenance from all duplicates
+
+**Configuration**:
+```toml
+[context.strategies]
+deduplicator_strategy = "highest_relevance"
+deduplicator_merge_provenance = true
+```
+
+**Example**:
+```python
+deduplicator = FragmentDeduplicator()
+deduplicated = deduplicator.deduplicate(
+    fragments=[frag1, frag2, frag1_duplicate],
+    strategy="highest_relevance"
+)
+# Returns: [frag1 (with merged provenance), frag2]
+```
+
+#### DetailDepthResolver
+
+**Purpose**: Resolve appropriate detail depth for each fragment.
+
+**Input**:
+- `fragments`: List of ContextFragment objects
+- `budget`: ContextBudget
+- `request`: ContextRequest with desired detail level
+
+**Output**:
+- `resolved`: List of fragments with adjusted detail_depth
+
+**Resolution Logic**:
+1. Start with requested detail level
+2. Adjust based on token budget
+3. Respect fragment's max depth
+4. Prefer higher depth for hot tier
+
+**Configuration**:
+```toml
+[context.strategies]
+resolver_default_depth = 3
+resolver_hot_tier_boost = 2
+resolver_budget_aware = true
+```
+
+**Example**:
+```python
+resolver = DetailDepthResolver()
+resolved = resolver.resolve(
+    fragments=[frag1, frag2],
+    budget=budget,
+    request=context_request
+)
+# Returns: fragments with adjusted detail_depth
+```
+
+#### FragmentScorer
+
+**Purpose**: Score fragments based on multiple factors.
+
+**Input**:
+- `fragments`: List of ContextFragment objects
+- `plan_context`: Current plan context
+- `request`: ContextRequest
+
+**Output**:
+- `scored`: List of (fragment, score) tuples
+
+**Scoring Factors**:
+- Relevance score (from strategy): 40%
+- Tier priority (hot > warm > cold): 30%
+- Recency (created_at): 20%
+- Plan context alignment: 10%
+
+**Configuration**:
+```toml
+[context.strategies]
+scorer_relevance_weight = 0.4
+scorer_tier_weight = 0.3
+scorer_recency_weight = 0.2
+scorer_plan_weight = 0.1
+```
+
+**Example**:
+```python
+scorer = FragmentScorer()
+scored = scorer.score(
+    fragments=[frag1, frag2],
+    plan_context=plan_context,
+    request=context_request
+)
+# Returns: [(frag1, 0.92), (frag2, 0.78)]
+```
+
+#### BudgetPacker
+
+**Purpose**: Select fragments to fit within token budget.
+
+**Input**:
+- `scored_fragments`: List of (fragment, score) tuples
+- `budget`: ContextBudget
+
+**Output**:
+- `packed`: List of selected fragments
+
+**Packing Algorithm**:
+1. Sort fragments by score (descending)
+2. Respect tier ordering (hot → warm → cold)
+3. Greedily add fragments until budget exhausted
+4. Early termination when budget full
+
+**Configuration**:
+```toml
+[context.strategies]
+packer_tier_priority = ["hot", "warm", "cold"]
+packer_greedy = true
+```
+
+**Example**:
+```python
+packer = BudgetPacker()
+packed = packer.pack(
+    scored_fragments=[(frag1, 0.92), (frag2, 0.78)],
+    budget=budget
+)
+# Returns: [frag1, frag2] (total tokens <= budget.available_tokens)
+```
+
+#### FragmentOrderer
+
+**Purpose**: Determine final fragment order in payload.
+
+**Input**:
+- `fragments`: List of ContextFragment objects
+- `strategy`: Ordering strategy ("relevance", "recency", "tier-relevance", "custom")
+
+**Output**:
+- `ordered`: List of fragments in final order
+
+**Ordering Strategies**:
+- **Relevance**: By relevance_score (descending)
+- **Recency**: By created_at (descending)
+- **Tier-Relevance**: By tier (hot > warm > cold), then relevance
+- **Custom**: User-defined ordering
+
+**Configuration**:
+```toml
+[context.strategies]
+orderer_strategy = "tier-relevance"
+```
+
+**Example**:
+```python
+orderer = FragmentOrderer()
+ordered = orderer.order(
+    fragments=[frag1, frag2],
+    strategy="tier-relevance"
+)
+# Returns: [hot_frag, warm_frag1, warm_frag2, cold_frag]
+```
+
+### Phase 3: Context Finalization
+
+#### PreambleGenerator
+
+**Purpose**: Generate optional summary of assembled context.
+
+**Input**:
+- `payload`: ContextPayload (before finalization)
+- `strategies_used`: Tuple of strategy names
+- `budget_used`: Fraction of budget consumed
+
+**Output**:
+- `preamble`: String summary of context
+
+**Preamble Contents**:
+- Number of fragments and sources
+- Token usage and budget utilization
+- Strategies used
+- Key topics covered
+- Temporal range
+
+**Configuration**:
+```toml
+[context.strategies]
+preamble_enabled = true
+preamble_include_stats = true
+preamble_max_length = 500
+```
+
+**Example**:
+```python
+generator = PreambleGenerator()
+preamble = generator.generate(
+    payload=payload,
+    strategies_used=["semantic-embedding"],
+    budget_used=0.57
+)
+# Returns: "Context includes 5 fragments from 3 sources..."
+```
+
+#### SkeletonCompressor
+
+**Purpose**: Compress context for child plans while preserving critical information.
+
+**Input**:
+- `parent_context`: Parent ContextPayload
+- `child_focus`: List of focus nodes for child plan
+- `skeleton_budget`: Token budget for compressed context
+
+**Output**:
+- `compressed`: Compressed ContextPayload for child plan
+
+**Compression Strategy**:
+1. Identify critical fragments for child focus
+2. Remove redundant details
+3. Preserve decision context
+4. Maintain provenance
+
+**Configuration**:
+```toml
+[context.strategies]
+compressor_enabled = true
+compressor_preserve_decisions = true
+compressor_preserve_hot_tier = true
+```
+
+**Example**:
+```python
+compressor = SkeletonCompressor()
+compressed = compressor.compress(
+    parent_context=parent_payload,
+    child_focus=["project://myapp/src/main.py"],
+    skeleton_budget=1024
+)
+# Returns: compressed ContextPayload for child plan
+```
+
+## Data Flow
+
+### Complete Pipeline Flow
+
+```
+ContextRequest
+    ↓
+StrategySelector
+    ↓ (candidates: [(strategy, confidence), ...])
+BudgetAllocator
+    ↓ (allocations: {strategy: budget, ...})
+StrategyExecutor
+    ↓ (results: [StrategyResult, ...])
+Merge Results
+    ↓ (fragments: [ContextFragment, ...])
+FragmentDeduplicator
+    ↓ (deduplicated: [ContextFragment, ...])
+DetailDepthResolver
+    ↓ (resolved: [ContextFragment, ...])
+FragmentScorer
+    ↓ (scored: [(ContextFragment, score), ...])
+BudgetPacker
+    ↓ (packed: [ContextFragment, ...])
+FragmentOrderer
+    ↓ (ordered: [ContextFragment, ...])
+PreambleGenerator
+    ↓ (preamble: str)
+SkeletonCompressor (optional)
+    ↓ (compressed: ContextPayload)
+ContextPayload
+```
+
+### Example Data Structures
+
+**ContextRequest**:
+```python
+{
+    "focus_nodes": ["project://myapp/src/main.py"],
+    "breadth": 2,
+    "depth": 3,
+    "keywords": ["async", "io"],
+    "resource_types": ["source_file", "decision"],
+    "detail_level": 3,
+    "temporal_window": "7d",
+    "strategy_hints": ["semantic-embedding"]
+}
+```
+
+**StrategyResult**:
+```python
+{
+    "strategy_name": "semantic-embedding",
+    "fragments": [
+        ContextFragment(...),
+        ContextFragment(...),
+        ...
+    ],
+    "total_fragments": 45,
+    "tokens_used": 2048,
+    "execution_time_ms": 234,
+    "errors": [],
+    "stats": {
+        "similarity_scores": [0.95, 0.87, ...],
+        "backend_queries": 3,
+        "cache_hits": 2
+    }
+}
+```
+
+**ContextPayload**:
+```python
+{
+    "payload_id": "01ARZ3NDEKTSV4RRFFQ69G5FAV",
+    "plan_id": "plan-1",
+    "fragments": [
+        ContextFragment(...),
+        ContextFragment(...),
+        ...
+    ],
+    "total_tokens": 2048,
+    "budget": {
+        "max_tokens": 4096,
+        "reserved_tokens": 512,
+        "available_tokens": 3584
+    },
+    "budget_used": 0.57,
+    "strategies_used": ["semantic-embedding"],
+    "context_hash": "abc123...",
+    "preamble": "Context includes 5 fragments...",
+    "provenance_map": {
+        "fragment-id-1": {...},
+        "fragment-id-2": {...},
+        ...
+    },
+    "assembled_at": "2024-04-19T14:30:00Z"
+}
+```
+
+## Configuration and Tuning
+
+### Global Configuration
+
+```toml
+[context]
+max_tokens = 4096
+reserved_tokens = 512
+
+[context.strategies]
+enabled = ["simple-keyword", "semantic-embedding", "breadth-depth-navigator"]
+default_strategy = "arce"
+
+[context.strategies.simple-keyword]
+enabled = true
+timeout_seconds = 10
+max_fragments = 50
+
+[context.strategies.semantic-embedding]
+enabled = true
+timeout_seconds = 20
+max_fragments = 100
+
+[context.strategies.breadth-depth-navigator]
+enabled = true
+timeout_seconds = 30
+max_fragments = 150
+
+[context.strategies.arce]
+enabled = true
+timeout_seconds = 45
+max_fragments = 200
+text_budget_fraction = 0.4
+vector_budget_fraction = 0.4
+graph_budget_fraction = 0.2
+```
+
+### Per-Component Tuning
+
+#### StrategySelector Tuning
+
+```toml
+[context.strategies]
+selector_threshold = 0.0      # Minimum confidence to select
+max_candidates = 10           # Maximum strategies to execute
+```
+
+**Tuning Tips**:
+- Increase `selector_threshold` to filter low-confidence strategies
+- Decrease `max_candidates` to reduce execution time
+
+#### BudgetAllocator Tuning
+
+```toml
+[context.strategies]
+allocator_strategy = "proportional"  # "proportional", "equal", "priority"
+allocator_min_budget = 100
+```
+
+**Tuning Tips**:
+- Use "proportional" for quality-aware allocation
+- Use "equal" for fair distribution
+- Use "priority" for greedy allocation
+
+#### StrategyExecutor Tuning
+
+```toml
+[context.strategies]
+executor_max_workers = 4
+executor_timeout_seconds = 30
+executor_circuit_breaker_threshold = 3
+executor_circuit_breaker_reset_timeout = 60
+```
+
+**Tuning Tips**:
+- Increase `max_workers` for parallel execution (up to CPU cores)
+- Increase `timeout_seconds` for complex strategies
+- Adjust `circuit_breaker_threshold` for fault tolerance
+
+#### FragmentScorer Tuning
+
+```toml
+[context.strategies]
+scorer_relevance_weight = 0.4
+scorer_tier_weight = 0.3
+scorer_recency_weight = 0.2
+scorer_plan_weight = 0.1
+```
+
+**Tuning Tips**:
+- Increase `relevance_weight` for quality-focused assembly
+- Increase `tier_weight` for tier-aware assembly
+- Increase `recency_weight` for recent-focused assembly
+
+## Monitoring and Metrics
+
+### Key Metrics
+
+| Metric | Description | Target |
+|--------|-------------|--------|
+| `assembly_time_ms` | Total pipeline execution time | < 1000ms |
+| `fragments_selected` | Number of fragments in payload | 5-50 |
+| `tokens_used` | Total tokens in payload | < available_tokens |
+| `budget_used` | Fraction of budget consumed | 0.5-0.9 |
+| `avg_relevance` | Average relevance score | > 0.7 |
+| `strategy_success_rate` | Fraction of successful strategies | > 0.8 |
+| `dedup_ratio` | Fraction of duplicates removed | 0.0-0.3 |
+
+### Logging
+
+Enable debug logging:
+
+```python
+import logging
+
+logging.getLogger("cleveragents.acms").setLevel(logging.DEBUG)
+```
+
+### Metrics Collection
+
+```python
+from cleveragents.application.services.acms_service import ACMSPipeline
+
+pipeline = ACMSPipeline()
+payload = pipeline.assemble(plan_id, fragments, budget)
+
+# Collect metrics
+metrics = {
+    "payload_id": payload.payload_id,
+    "plan_id": payload.plan_id,
+    "total_tokens": payload.total_tokens,
+    "budget_used": payload.budget_used,
+    "fragments_count": len(payload.fragments),
+    "strategies_used": payload.strategies_used,
+    "avg_relevance": sum(f.relevance_score for f in payload.fragments) / len(payload.fragments),
+    "tier_distribution": {
+        "hot": sum(1 for f in payload.fragments if f.tier == "hot"),
+        "warm": sum(1 for f in payload.fragments if f.tier == "warm"),
+        "cold": sum(1 for f in payload.fragments if f.tier == "cold"),
+    }
+}
+```
+
+## Performance Optimization
+
+### Strategy Selection Optimization
+
+```python
+# Good: Filter low-confidence strategies
+candidates = [
+    (s, conf) for s, conf in candidates if conf >= 0.7
+]
+
+# Avoid: Execute all strategies
+candidates = all_strategies
+```
+
+### Budget Allocation Optimization
+
+```python
+# Good: Proportional allocation
+allocations = allocator.allocate(
+    candidates=high_confidence,
+    total_budget=3584,
+    strategy="proportional"
+)
+
+# Avoid: Equal allocation
+allocations = {s: 3584 // len(strategies) for s in strategies}
+```
+
+### Execution Optimization
+
+```python
+# Good: Parallel execution with reasonable timeout
+executor = StrategyExecutor(max_workers=4, timeout_seconds=30)
+
+# Avoid: Sequential execution or excessive timeout
+executor = StrategyExecutor(max_workers=1, timeout_seconds=300)
+```
+
+### Fragment Deduplication Optimization
+
+```python
+# Good: Deduplicate early
+deduplicated = deduplicator.deduplicate(fragments)
+scored = scorer.score(deduplicated, plan_context)
+
+# Avoid: Deduplicate after scoring
+scored = scorer.score(fragments, plan_context)
+deduplicated = deduplicator.deduplicate(scored)
+```
+
+## Troubleshooting
+
+### Slow Pipeline Execution
+
+**Problem**: Pipeline takes > 1 second.
+
+**Solution**:
+1. Check `executor_timeout_seconds` - reduce if too high
+2. Check `executor_max_workers` - increase if too low
+3. Check strategy complexity - use simpler strategies
+4. Check fragment count - reduce `max_fragments`
+
+### Low Quality Context
+
+**Problem**: Assembled context is not relevant.
+
+**Solution**:
+1. Increase `selector_threshold` to filter low-confidence strategies
+2. Use higher-quality strategy (e.g., ARCE)
+3. Increase `scorer_relevance_weight` for quality focus
+4. Adjust strategy-specific parameters
+
+### Out of Memory
+
+**Problem**: Pipeline consumes too much memory.
+
+**Solution**:
+1. Reduce `max_fragments` limit
+2. Reduce `executor_max_workers` for parallel execution
+3. Enable fragment streaming
+4. Use simpler strategies
+
+### Budget Exceeded
+
+**Problem**: Assembled context exceeds budget.
+
+**Solution**:
+1. Check `BudgetPacker` implementation
+2. Reduce `max_fragments` limit
+3. Increase `reserved_tokens` to reduce available budget
+4. Enable detail depth reduction
+
+## See Also
+
+- [Context Management Deep Dive](../guides/context_management_deep_dive.md)
+- [Context Strategies Reference](./context_strategies_reference.md)
+- [ACMS v1 Context Assembly Pipeline](./acms.md)
+- [Context Strategy Registry](./context_strategies.md)

docs/reference/context_strategies_reference.md

@@ -0,0 +1,750 @@
+# Context Strategies Reference
+
+Complete reference for all context strategies in the ACMS, including configuration options, performance characteristics, and use case recommendations.
+
+## Strategy Overview
+
+| Strategy | Quality | Speed | Coverage | Backends | Best For |
+|----------|---------|-------|----------|----------|----------|
+| Simple Keyword | 0.3 | ⚡⚡⚡ | Narrow | Text | Quick searches, fallback |
+| Semantic Embedding | 0.6 | ⚡⚡ | Broad | Vector | Semantic similarity |
+| Breadth-Depth Navigator | 0.85 | ⚡ | Very Broad | Graph | Comprehensive context |
+| ARCE | 0.95 | 🐢 | Comprehensive | All | Production use |
+| Temporal Archaeology | 0.5 | 🐢 | Historical | Graph + Temporal | Historical context |
+| Plan-Decision Context | 0.7 | ⚡ | Hierarchical | Temporal + Graph | Plan coordination |
+
+## Simple Keyword Strategy
+
+### Overview
+
+The Simple Keyword Strategy performs text-based keyword search using the TextBackend. It's the fastest strategy and serves as a fallback when other backends are unavailable.
+
+### Quality Score
+
+**0.3** (baseline)
+
+### Backend Requirements
+
+- **Required**: TextBackend
+- **Optional**: None
+
+### How It Works
+
+1. Extract keywords from the ContextRequest
+2. Query TextBackend using keyword search
+3. Rank results by match frequency and relevance
+4. Pack greedily into token budget
+
+### Configuration
+
+```toml
+[context.strategies.simple-keyword]
+enabled = true
+timeout_seconds = 10
+max_fragments = 50
+keyword_boost_factor = 1.5
+```
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `enabled` | bool | true | Enable this strategy |
+| `timeout_seconds` | int | 10 | Per-strategy timeout |
+| `max_fragments` | int | 50 | Maximum fragments returned |
+| `keyword_boost_factor` | float | 1.5 | Boost for exact keyword matches |
+
+### Performance Characteristics
+
+- **Execution Time**: 50-200ms (typical)
+- **Memory Usage**: Low (streaming results)
+- **Scalability**: Excellent (linear with keyword count)
+- **Parallelization**: Not applicable (single-threaded)
+
+### Use Cases
+
+- Quick searches with clear keywords
+- Projects with good text indexing
+- Fallback when other backends unavailable
+- Real-time context assembly with tight latency budgets
+
+### Example Usage
+
+```python
+from cleveragents.application.services.acms_service import ACMSPipeline
+
+pipeline = ACMSPipeline()
+payload = pipeline.assemble(
+    plan_id="plan-1",
+    fragments=fragments,
+    budget=ContextBudget(max_tokens=4096, reserved_tokens=512),
+    strategy="simple-keyword"
+)
+```
+
+### Limitations
+
+- Limited to exact keyword matches
+- No semantic understanding
+- Cannot traverse relationships
+- Requires good text indexing
+
+## Semantic Embedding Strategy
+
+### Overview
+
+The Semantic Embedding Strategy uses vector similarity search to find semantically related context. It uses character-frequency embeddings as a v1 approximation.
+
+### Quality Score
+
+**0.6** (good)
+
+### Backend Requirements
+
+- **Required**: VectorBackend
+- **Optional**: None
+
+### How It Works
+
+1. Embed the ContextRequest using character-frequency model
+2. Query VectorBackend for similar embeddings
+3. Rank by cosine similarity
+4. Pack budget-aware results
+
+### Configuration
+
+```toml
+[context.strategies.semantic-embedding]
+enabled = true
+timeout_seconds = 20
+max_fragments = 100
+similarity_threshold = 0.5
+embedding_model = "character-frequency-v1"
+```
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `enabled` | bool | true | Enable this strategy |
+| `timeout_seconds` | int | 20 | Per-strategy timeout |
+| `max_fragments` | int | 100 | Maximum fragments returned |
+| `similarity_threshold` | float | 0.5 | Minimum similarity score (0.0-1.0) |
+| `embedding_model` | str | "character-frequency-v1" | Embedding model to use |
+
+### Performance Characteristics
+
+- **Execution Time**: 100-500ms (typical)
+- **Memory Usage**: Medium (embedding cache)
+- **Scalability**: Good (logarithmic with vector index size)
+- **Parallelization**: Supported (multiple similarity searches)
+
+### Use Cases
+
+- Semantic similarity searches
+- Conceptual relationships
+- Cross-domain context discovery
+- Finding related code patterns
+
+### Example Usage
+
+```python
+from cleveragents.domain.models.core.context_fragment import ContextRequest
+
+request = ContextRequest(
+    focus_nodes=["project://myapp/src/main.py"],
+    keywords=["async", "io"]
+)
+
+payload = pipeline.assemble(
+    plan_id="plan-1",
+    fragments=fragments,
+    budget=ContextBudget(max_tokens=4096, reserved_tokens=512),
+    strategy="semantic-embedding"
+)
+```
+
+### Limitations
+
+- v1 uses character-frequency embeddings (not true semantic)
+- Requires VectorBackend
+- Cannot traverse relationships
+- Sensitive to embedding quality
+
+### Future Improvements
+
+- Integration with real embedding models (e.g., OpenAI, Anthropic)
+- Fine-tuned embeddings for code
+- Hybrid semantic-keyword search
+
+## Breadth-Depth Navigator Strategy
+
+### Overview
+
+The Breadth-Depth Navigator Strategy traverses the knowledge graph to find related context. It explores outward (breadth) and downward (depth) from focus nodes.
+
+### Quality Score
+
+**0.85** (excellent)
+
+### Backend Requirements
+
+- **Required**: GraphBackend
+- **Optional**: None
+
+### How It Works
+
+1. Start from focus nodes in ContextRequest
+2. Traverse graph outward by breadth hops
+3. Traverse downward by depth hops
+4. Collect all reachable nodes
+5. Rank by distance and relevance
+6. Pack into budget
+
+### Configuration
+
+```toml
+[context.strategies.breadth-depth-navigator]
+enabled = true
+timeout_seconds = 30
+max_fragments = 150
+default_breadth = 2
+default_depth = 3
+distance_decay = 0.8
+```
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `enabled` | bool | true | Enable this strategy |
+| `timeout_seconds` | int | 30 | Per-strategy timeout |
+| `max_fragments` | int | 150 | Maximum fragments returned |
+| `default_breadth` | int | 2 | Default breadth hops |
+| `default_depth` | int | 3 | Default depth hops |
+| `distance_decay` | float | 0.8 | Decay factor per hop |
+
+### Performance Characteristics
+
+- **Execution Time**: 200-1000ms (typical)
+- **Memory Usage**: Medium (graph traversal state)
+- **Scalability**: Good (exponential with breadth/depth)
+- **Parallelization**: Supported (parallel hop traversal)
+
+### Use Cases
+
+- Exploring related code and decisions
+- Understanding dependencies
+- Comprehensive context assembly
+- Finding transitive relationships
+
+### Example Usage
+
+```python
+request = ContextRequest(
+    focus_nodes=["project://myapp/src/main.py"],
+    breadth=2,
+    depth=3
+)
+
+payload = pipeline.assemble(
+    plan_id="plan-1",
+    fragments=fragments,
+    budget=ContextBudget(max_tokens=4096, reserved_tokens=512),
+    strategy="breadth-depth-navigator"
+)
+```
+
+### Limitations
+
+- Requires GraphBackend
+- Can be slow with large graphs
+- Exponential complexity with breadth/depth
+- May include irrelevant distant nodes
+
+### Optimization Tips
+
+- Reduce breadth/depth for faster execution
+- Use distance decay to prioritize nearby nodes
+- Filter by resource type to reduce graph size
+- Increase timeout for large graphs
+
+## ARCE Strategy (Advanced Retrieval with Context Enrichment)
+
+### Overview
+
+ARCE is the most comprehensive strategy, combining text search, vector similarity, and graph traversal. It's the recommended strategy for production use cases.
+
+### Quality Score
+
+**0.95** (highest)
+
+### Backend Requirements
+
+- **Required**: TextBackend, VectorBackend, GraphBackend
+- **Optional**: None
+
+### How It Works
+
+1. **Phase 1 (40% budget)**: Text search using keywords
+2. **Phase 2 (40% budget)**: Vector similarity search
+3. **Phase 3 (20% budget)**: Graph traversal from results
+4. Merge and deduplicate results
+5. Rank by combined score
+6. Pack into budget
+
+### Configuration
+
+```toml
+[context.strategies.arce]
+enabled = true
+timeout_seconds = 45
+max_fragments = 200
+text_budget_fraction = 0.4
+vector_budget_fraction = 0.4
+graph_budget_fraction = 0.2
+dedup_strategy = "highest_relevance"
+```
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `enabled` | bool | true | Enable this strategy |
+| `timeout_seconds` | int | 45 | Per-strategy timeout |
+| `max_fragments` | int | 200 | Maximum fragments returned |
+| `text_budget_fraction` | float | 0.4 | Budget for text search |
+| `vector_budget_fraction` | float | 0.4 | Budget for vector search |
+| `graph_budget_fraction` | float | 0.2 | Budget for graph traversal |
+| `dedup_strategy` | str | "highest_relevance" | Deduplication strategy |
+
+### Performance Characteristics
+
+- **Execution Time**: 500-2000ms (typical)
+- **Memory Usage**: High (multiple result sets)
+- **Scalability**: Good (parallel phases)
+- **Parallelization**: Excellent (3 phases in parallel)
+
+### Use Cases
+
+- Comprehensive context assembly
+- Multi-modal queries
+- Production use cases
+- High-quality context requirements
+
+### Example Usage
+
+```python
+request = ContextRequest(
+    focus_nodes=["project://myapp/src/main.py"],
+    keywords=["async", "io"],
+    breadth=2
+)
+
+payload = pipeline.assemble(
+    plan_id="plan-1",
+    fragments=fragments,
+    budget=ContextBudget(max_tokens=4096, reserved_tokens=512),
+    strategy="arce"
+)
+```
+
+### Limitations
+
+- Requires all three backends
+- Slower than single-backend strategies
+- Higher memory usage
+- May return too many fragments
+
+### Optimization Tips
+
+- Adjust budget fractions based on query type
+- Increase timeout for complex queries
+- Use deduplication to reduce redundancy
+- Filter by resource type to reduce results
+
+## Temporal Archaeology Strategy
+
+### Overview
+
+The Temporal Archaeology Strategy retrieves historical context from the cold tier and traverses relationships from historical nodes.
+
+### Quality Score
+
+**0.5** (moderate)
+
+### Backend Requirements
+
+- **Required**: GraphBackend, TemporalBackend (cold tier)
+- **Optional**: None
+
+### How It Works
+
+1. Query TemporalBackend for historical nodes
+2. Filter by temporal window
+3. Traverse graph from historical nodes
+4. Collect related current nodes
+5. Rank by recency and relevance
+6. Pack into budget
+
+### Configuration
+
+```toml
+[context.strategies.temporal-archaeology]
+enabled = false
+timeout_seconds = 60
+max_fragments = 100
+default_temporal_window = "7d"
+include_current_context = true
+```
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `enabled` | bool | false | Enable this strategy |
+| `timeout_seconds` | int | 60 | Per-strategy timeout |
+| `max_fragments` | int | 100 | Maximum fragments returned |
+| `default_temporal_window` | str | "7d" | Default time window |
+| `include_current_context` | bool | true | Include current context |
+
+### Performance Characteristics
+
+- **Execution Time**: 1000-5000ms (typical)
+- **Memory Usage**: High (temporal queries)
+- **Scalability**: Fair (depends on cold tier size)
+- **Parallelization**: Limited (sequential phases)
+
+### Use Cases
+
+- Understanding historical context
+- Decision archaeology
+- Long-term project evolution
+- Regression analysis
+
+### Example Usage
+
+```python
+request = ContextRequest(
+    focus_nodes=["project://myapp/src/main.py"],
+    temporal_window="30d"
+)
+
+payload = pipeline.assemble(
+    plan_id="plan-1",
+    fragments=fragments,
+    budget=ContextBudget(max_tokens=4096, reserved_tokens=512),
+    strategy="temporal-archaeology"
+)
+```
+
+### Limitations
+
+- Requires cold tier (not always available)
+- Slow execution
+- May include outdated information
+- Limited to historical data
+
+### Temporal Window Format
+
+- `"1d"` - Last 1 day
+- `"7d"` - Last 7 days
+- `"30d"` - Last 30 days
+- `"90d"` - Last 90 days
+- `"1y"` - Last 1 year
+- `"all"` - All history
+
+## Plan-Decision Context Strategy
+
+### Overview
+
+The Plan-Decision Context Strategy walks the plan hierarchy to retrieve decision records and related context.
+
+### Quality Score
+
+**0.7** (good)
+
+### Backend Requirements
+
+- **Required**: TemporalBackend (warm/cold), GraphBackend
+- **Optional**: None
+
+### How It Works
+
+1. Walk parent and ancestor plan hierarchy
+2. Retrieve decision records from warm/cold tiers
+3. Traverse graph from decisions
+4. Collect related context
+5. Rank by plan proximity and relevance
+6. Pack into budget
+
+### Configuration
+
+```toml
+[context.strategies.plan-decision-context]
+enabled = true
+timeout_seconds = 30
+max_fragments = 100
+include_ancestor_decisions = true
+max_ancestor_depth = 5
+```
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `enabled` | bool | true | Enable this strategy |
+| `timeout_seconds` | int | 30 | Per-strategy timeout |
+| `max_fragments` | int | 100 | Maximum fragments returned |
+| `include_ancestor_decisions` | bool | true | Include ancestor decisions |
+| `max_ancestor_depth` | int | 5 | Maximum ancestor levels |
+
+### Performance Characteristics
+
+- **Execution Time**: 200-800ms (typical)
+- **Memory Usage**: Medium (plan hierarchy)
+- **Scalability**: Good (linear with plan depth)
+- **Parallelization**: Supported (parallel ancestor traversal)
+
+### Use Cases
+
+- Understanding plan context
+- Decision inheritance
+- Multi-level plan coordination
+- Plan dependency analysis
+
+### Example Usage
+
+```python
+request = ContextRequest(
+    focus_nodes=["project://myapp/src/main.py"]
+)
+
+payload = pipeline.assemble(
+    plan_id="plan-1",
+    fragments=fragments,
+    budget=ContextBudget(max_tokens=4096, reserved_tokens=512),
+    strategy="plan-decision-context"
+)
+```
+
+### Limitations
+
+- Requires plan hierarchy
+- Limited to decision context
+- May miss non-decision context
+- Depends on plan structure
+
+## Strategy Selection Guide
+
+### When to Use Each Strategy
+
+#### Simple Keyword
+- ✅ Quick searches with clear keywords
+- ✅ Fallback when other backends unavailable
+- ✅ Real-time context with tight latency budgets
+- ❌ Semantic understanding needed
+- ❌ Relationship traversal needed
+
+#### Semantic Embedding
+- ✅ Semantic similarity searches
+- ✅ Conceptual relationships
+- ✅ Cross-domain discovery
+- ❌ Exact keyword matching needed
+- ❌ Relationship traversal needed
+
+#### Breadth-Depth Navigator
+- ✅ Exploring related code
+- ✅ Understanding dependencies
+- ✅ Comprehensive context
+- ❌ Quick searches needed
+- ❌ Historical context needed
+
+#### ARCE
+- ✅ Comprehensive context assembly
+- ✅ Multi-modal queries
+- ✅ Production use cases
+- ❌ Quick searches needed
+- ❌ Limited backends available
+
+#### Temporal Archaeology
+- ✅ Historical context
+- ✅ Decision archaeology
+- ✅ Long-term evolution
+- ❌ Current context needed
+- ❌ Quick searches needed
+
+#### Plan-Decision Context
+- ✅ Plan coordination
+- ✅ Decision inheritance
+- ✅ Multi-level plans
+- ❌ Non-decision context needed
+- ❌ No plan hierarchy
+
+### Fallback Chain
+
+When a strategy fails, the system falls back to simpler strategies:
+
+```
+ARCE (0.95)
+  ↓ (if unavailable)
+Breadth-Depth Navigator (0.85)
+  ↓ (if unavailable)
+Semantic Embedding (0.6)
+  ↓ (if unavailable)
+Simple Keyword (0.3)
+```
+
+## Performance Comparison
+
+### Execution Time
+
+```
+Simple Keyword:        ████ 50-200ms
+Semantic Embedding:    ██████████ 100-500ms
+Breadth-Depth:         ███████████████ 200-1000ms
+Plan-Decision:         ██████████ 200-800ms
+ARCE:                  ████████████████████ 500-2000ms
+Temporal Archaeology:  ████████████████████████ 1000-5000ms
+```
+
+### Memory Usage
+
+```
+Simple Keyword:        ██ Low
+Semantic Embedding:    ████ Medium
+Breadth-Depth:         ████ Medium
+Plan-Decision:         ████ Medium
+ARCE:                  ██████ High
+Temporal Archaeology:  ██████ High
+```
+
+### Coverage
+
+```
+Simple Keyword:        ██ Narrow
+Temporal Archaeology:  ███ Moderate
+Semantic Embedding:    ████ Broad
+Plan-Decision:         ████ Hierarchical
+Breadth-Depth:         █████ Very Broad
+ARCE:                  ██████ Comprehensive
+```
+
+## Configuration Best Practices
+
+### Global Configuration
+
+```toml
+[context.strategies]
+enabled = ["simple-keyword", "semantic-embedding", "breadth-depth-navigator"]
+default_strategy = "arce"
+fallback_chain = ["arce", "breadth-depth-navigator", "semantic-embedding", "simple-keyword"]
+
+[context.strategies.simple-keyword]
+enabled = true
+timeout_seconds = 10
+max_fragments = 50
+
+[context.strategies.semantic-embedding]
+enabled = true
+timeout_seconds = 20
+max_fragments = 100
+
+[context.strategies.breadth-depth-navigator]
+enabled = true
+timeout_seconds = 30
+max_fragments = 150
+
+[context.strategies.arce]
+enabled = true
+timeout_seconds = 45
+max_fragments = 200
+
+[context.strategies.plan-decision-context]
+enabled = true
+timeout_seconds = 30
+max_fragments = 100
+```
+
+### Per-Project Overrides
+
+```bash
+# Set strategies for a project
+agents project context set \
+  --strategy simple-keyword \
+  --strategy semantic-embedding \
+  --strategy breadth-depth-navigator
+
+# View current strategies
+agents project context list
+```
+
+## Monitoring and Metrics
+
+### Key Metrics
+
+- **Execution Time**: Strategy execution wall-clock time
+- **Fragment Count**: Number of fragments returned
+- **Token Usage**: Total tokens in assembled context
+- **Budget Utilization**: Percentage of budget used
+- **Relevance Score**: Average relevance of fragments
+- **Tier Distribution**: Hot/warm/cold fragment counts
+
+### Logging
+
+Enable debug logging to monitor strategy execution:
+
+```python
+import logging
+
+logging.getLogger("cleveragents.acms").setLevel(logging.DEBUG)
+```
+
+### Example Metrics Output
+
+```
+Strategy: semantic-embedding
+  Execution Time: 234ms
+  Fragments: 12
+  Tokens: 2048
+  Budget Used: 57%
+  Avg Relevance: 0.82
+  Tier Distribution: hot=2, warm=8, cold=2
+```
+
+## Troubleshooting
+
+### Strategy Not Selected
+
+**Problem**: Expected strategy not selected.
+
+**Solution**:
+1. Check strategy is enabled in configuration
+2. Verify `can_handle()` returns confidence > 0
+3. Check required backends are available
+4. Review strategy selection logs
+
+### Slow Execution
+
+**Problem**: Strategy takes too long.
+
+**Solution**:
+1. Reduce `max_fragments` limit
+2. Decrease `breadth` and `depth` parameters
+3. Increase `timeout_seconds` if appropriate
+4. Switch to faster strategy
+
+### Low Quality Results
+
+**Problem**: Assembled context is not relevant.
+
+**Solution**:
+1. Try higher-quality strategy (e.g., ARCE)
+2. Adjust `similarity_threshold` for semantic strategy
+3. Increase `breadth` and `depth` for navigator
+4. Provide better keywords in request
+
+### Out of Memory
+
+**Problem**: Strategy consumes too much memory.
+
+**Solution**:
+1. Reduce `max_fragments` limit
+2. Switch to lower-memory strategy
+3. Reduce `breadth` and `depth` parameters
+4. Enable fragment streaming
+
+## See Also
+
+- [Context Management Deep Dive](../guides/context_management_deep_dive.md)
+- [ACMS Pipeline Reference](./acms_pipeline_reference.md)
+- [ACMS v1 Context Assembly Pipeline](./acms.md)
+- [Context Strategy Registry](./context_strategies.md)

mkdocs.yml

@@ -24,8 +24,6 @@ nav:
     - AI Providers: api/providers.md
     - TUI: api/tui.md
     - ACMS / UKO: api/acms.md
-  - Guides:
-    - Configuration Reference: guides/configuration-reference.md
   - Modules:
     - Shell Safety: modules/shell-safety.md
     - UKO Provenance Tracking: modules/uko-provenance.md