Context Management

Context Management Skill

Unified context-window management with two distinct lifecycles:

• Runtime Compression — what an agent does before loading a large artifact to stay under the model context limit (during workflow execution).
• Diagnostic Audit — what the 11-Context Optimizer agent does after the fact to find waste in agent definitions, instructions, and skill loads.

The two modes do not depend on each other — pick the section that matches your need.

Mode A: Runtime Compression

Replaces the legacy context-shredding skill.

When to Use Runtime Compression

• Before loading a predecessor artifact file (01 through 07)
• When conversation length suggests >60% of model context is used
• When an agent needs to load multiple large artifacts

Compression Tiers

Hard Token Checkpoints

Percentages are advisory; absolute input-token counts override them. GPT-5.6-Terra and Luna hard-checkpoints at ≥300K input; Claude Opus 5 at ≥160K. When hit, emit a compaction message and switch every further read to the minimal tier. Full per-model table, checkpoint procedure (4 steps), and background context (nordic-foods saturation event) in references/hard-checkpoints.md.

Rules

1. Estimate context usage — count approximate conversation tokens
2. Select tier based on the thresholds above
3. Apply compression template from references/compression-templates.md
4. If loading multiple artifacts, compress the older / less-critical ones first

Steps

1. Estimate current context usage (rough: 1 token ≈ 4 chars)
2. Check model limit (Claude family: 200K, GPT-5 family: 400K)
3. Calculate usage percentage and check hard-checkpoint table
4. Select tier:
   < 60%  → full (no compression needed)
   60-80% → summarized (key sections only)
   > 80%  → minimal (decision summaries only)
5. Load artifact/skill using the appropriate variant

Skill Loading

Skills are single-tier — one file per skill, no digest / minimal variants. Load each SKILL.md only once per session; defer references/*.md until the SKILL.md body explicitly points to one. Full protocol in references/skill-loading.md.

Mode B: Diagnostic Audit

Replaces the legacy context-optimizer skill.

Structured methodology for auditing how GitHub Copilot agents consume their context window. Identifies waste, recommends hand-off points, and produces prioritised optimisation reports.

When to Use Diagnostic Audit

• Auditing context-window efficiency across a multi-agent system
• Identifying where to introduce subagent hand-offs
• Reducing redundant file reads and skill loads
• Optimising instruction file applyTo glob patterns
• Profiling per-turn token cost from debug logs
• Porting agent optimisations to a new project

Audit Capabilities & Prerequisites

Capabilities cover log parsing, turn-cost profiling, redundancy detection, hand-off gap analysis, instruction audit, and structured report generation. Prerequisites: Python 3.14, VS Code Copilot Chat debug logs, and .github/agents/*.agent.md (or equivalent). Full capability matrix, portability checklist, and debug-log discovery in references/audit-setup.md.

Analysis Methodology

For the complete methodology — log format reference (ccreq line parsing, request types, latency heuristics), Steps 1-5 (log parsing → optimisation recommendations), common optimisation patterns, and baseline comparison workflow (Phase 0 + Phase 6) — read references/analysis-methodology.md.

Report Template

See templates/optimization-report.md for the full output template.

Reference Index

Load on demand: