Context Window Optimization

TL;DR: Optimizes context window usage through progressive MOAT loading (L1/L2/L3), intelligent pruning, session state management, and token-efficient skill routing. Ensures the AI agent operates within context limits while maintaining access to the knowledge needed for the current task.

Principio Rector

El contexto es un recurso finito. Cargar los 100 skills completos excede cualquier ventana de contexto. La carga progresiva (L1 metadata, L2 core, L3 deep) permite acceder al conocimiento correcto en el momento correcto. La optimización de contexto no es ahorro — es precisión en la información cargada.

Assumptions & Limits

• Assumes skill catalog is indexed with loading level metadata (L1/L2/L3) [PLAN]
• Assumes the lazy-load-resolver script is functional and up to date [SUPUESTO]
• Breaks when context window is too small for even L1 metadata of required skills
• Does not optimize user-provided content — only framework-loaded content
• Session state persistence depends on project/ directory writability [SUPUESTO]
• Pruning decisions are heuristic — may occasionally remove still-relevant context

Usage

# Optimize context for a specific phase and project type
/pm:context-optimization $PROJECT --phase="planning" --tipo="agile"

# Analyze current context usage and recommend pruning
/pm:context-optimization $PROJECT --type=analyze

# Configure session state persistence rules
/pm:context-optimization $PROJECT --type=session-state --persist="essential"

Parameters:

Service Type Routing

{TIPO_PROYECTO}: All project types benefit from context optimization. Complex engagements need L3 for active skills; routine operations use L1/L2.

Before Optimizing

1. Read the current session state to understand what is already loaded
2. Read the skill catalog to determine which skills are relevant to the current task
3. Glob scripts/lazy-load-resolver.sh to verify resolver availability
4. Grep for project/session-state.json to check current context configuration

Entrada (Input Requirements)

• Current task and phase
• Available context window size
• Skill catalog with loading levels
• Session state
• Previous context usage patterns

Proceso (Protocol)

1. Task analysis — Determine which skills are relevant to current task
2. Loading level selection — Choose L1 (metadata), L2 (core), or L3 (deep) per skill
3. Priority ordering — Load highest-priority skills first
4. Session state management — Maintain essential state across interactions
5. Pruning strategy — Remove context no longer relevant to current task
6. Lazy loading — Load additional context on-demand when referenced
7. Compression — Summarize verbose context into essential information
8. Cache strategy — Define what to keep in persistent session state
9. Monitor usage — Track context consumption per interaction
10. Optimization report — Report context efficiency metrics

Edge Cases

1. Context overflow despite optimization: Emergency pruning — keep only active skill L2 + session state. Archive other context to project/context-archive/. Notify user of reduced capability. [PLAN]
2. Multiple skills needed simultaneously: Load all at L1 first. Promote to L2 on demand. Only one skill at L3 at a time. Document which skill is in focus. [METRIC]
3. Session state lost between interactions: Rebuild from project/session-state.json. If file missing, re-prime from last known good state. Flag data loss to user. [SUPUESTO]
4. User provides massive input exceeding context budget: Summarize user input preserving key facts. Store full input in project/ for reference. Process in chunks if needed. [PLAN]

Example: Good vs Bad

Good Context Optimization:

Bad Context Optimization: Loading all 100 skills at L3 into context, overflowing the window, and producing degraded responses because critical information is truncated. No pruning, no prioritization, no session state management. Fails because it treats context as infinite rather than as a resource to be managed.

Validation Gate

• Every loaded skill has explicit loading level (L1/L2/L3) justified by task relevance
• Context utilization ≤85% of available window to leave room for user interaction
• Session state persisted to project/session-state.json after each significant interaction
• Pruning removes ≥1 irrelevant context element per optimization cycle
• No skill loaded at L3 unless it is the active focus of the current task
• Lazy loading triggers correctly when skills are referenced but not yet loaded
• Context efficiency improvement ≥20% vs. naive full-loading approach
• Missing context risks flagged when pruning removes potentially relevant information
• Agent performance maintained despite context constraints [STAKEHOLDER]
• Context strategy aligns with pipeline workflow and phase transitions [PLAN]

Escalation Triggers

• Critical skill context unavailable due to window limits
• Context overflow causing degraded responses
• Session state loss between interactions
• Loading strategy causing performance issues

Additional Resources

Output Configuration

• Language: Spanish (Latin American, business register)
• Evidence: [PLAN], [SCHEDULE], [METRIC], [INFERENCIA], [SUPUESTO], [STAKEHOLDER]
• Branding: #2563EB royal blue, #F59E0B amber (NEVER green), #0F172A dark

Sub-Agents

Context Pruner

Context Pruner Agent

Core Responsibility

Prunes stale or low-priority content from context. This agent operates autonomously, applying systematic analysis and producing structured outputs.

Process

1. Gather Inputs. Collect all relevant data, documents, and stakeholder inputs needed for analysis.
2. Analyze Context. Assess the project context, methodology, phase, and constraints.
3. Apply Framework. Apply the appropriate analytical framework or model.
4. Generate Findings. Produce detailed findings with evidence tags and quantified impacts.
5. Validate Results. Cross-check findings against related artifacts for consistency.
6. Formulate Recommendations. Transform findings into actionable recommendations with owners and timelines.
7. Deliver Output. Produce the final structured output with executive summary, analysis, and action items.

Output Format

• Analysis Report — Structured findings with evidence tags and severity ratings.
• Recommendation Register — Actionable items with owners, deadlines, and success criteria.
• Executive Summary — 3-5 bullet point summary for stakeholder communication.

Lazy Resolver

Lazy Resolver Agent

Core Responsibility

Resolves lazy-loaded content on demand. This agent operates autonomously, applying systematic analysis and producing structured outputs.

Process

1. Gather Inputs. Collect all relevant data, documents, and stakeholder inputs needed for analysis.
2. Analyze Context. Assess the project context, methodology, phase, and constraints.
3. Apply Framework. Apply the appropriate analytical framework or model.
4. Generate Findings. Produce detailed findings with evidence tags and quantified impacts.
5. Validate Results. Cross-check findings against related artifacts for consistency.
6. Formulate Recommendations. Transform findings into actionable recommendations with owners and timelines.
7. Deliver Output. Produce the final structured output with executive summary, analysis, and action items.

Output Format

• Analysis Report — Structured findings with evidence tags and severity ratings.
• Recommendation Register — Actionable items with owners, deadlines, and success criteria.
• Executive Summary — 3-5 bullet point summary for stakeholder communication.

Progressive Loader

Progressive Loader Agent

Core Responsibility

Implements progressive loading for skill and reference content. This agent operates autonomously, applying systematic analysis and producing structured outputs.

Process

1. Gather Inputs. Collect all relevant data, documents, and stakeholder inputs needed for analysis.
2. Analyze Context. Assess the project context, methodology, phase, and constraints.
3. Apply Framework. Apply the appropriate analytical framework or model.
4. Generate Findings. Produce detailed findings with evidence tags and quantified impacts.
5. Validate Results. Cross-check findings against related artifacts for consistency.
6. Formulate Recommendations. Transform findings into actionable recommendations with owners and timelines.
7. Deliver Output. Produce the final structured output with executive summary, analysis, and action items.

Output Format

• Analysis Report — Structured findings with evidence tags and severity ratings.
• Recommendation Register — Actionable items with owners, deadlines, and success criteria.
• Executive Summary — 3-5 bullet point summary for stakeholder communication.

Token Budget Analyzer

Token Budget Analyzer Agent

Core Responsibility

Analyzes context window token usage and optimization opportunities. This agent operates autonomously, applying systematic analysis and producing structured outputs.

Process

1. Gather Inputs. Collect all relevant data, documents, and stakeholder inputs needed for analysis.
2. Analyze Context. Assess the project context, methodology, phase, and constraints.
3. Apply Framework. Apply the appropriate analytical framework or model.
4. Generate Findings. Produce detailed findings with evidence tags and quantified impacts.
5. Validate Results. Cross-check findings against related artifacts for consistency.
6. Formulate Recommendations. Transform findings into actionable recommendations with owners and timelines.
7. Deliver Output. Produce the final structured output with executive summary, analysis, and action items.

Output Format

• Analysis Report — Structured findings with evidence tags and severity ratings.
• Recommendation Register — Actionable items with owners, deadlines, and success criteria.
• Executive Summary — 3-5 bullet point summary for stakeholder communication.