Context Window Optimization Skill

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

When to Use This Skill

• Auditing context window efficiency across a multi-agent system
• Identifying where to introduce subagent hand-offs
• Reducing redundant file reads and skill loads
• Optimizing instruction file

  applyTo

  glob patterns
• Profiling per-turn token cost from debug logs
• Porting agent optimizations to a new project

Quick Reference

Prerequisites

• Python 3.10+ (for log parser script)
• Access to VS Code Copilot Chat debug logs
• Agent definitions in

  .github/agents/*.agent.md

  (or equivalent)

Enabling Debug Logs

Copilot Chat writes debug logs automatically to the VS Code log directory. To find the latest logs:

find ~/.vscode-server/data/logs/ -name "GitHub Copilot Chat.log" -newer /tmp/marker 2>/dev/null \
  | sort | tail -5

For richer output, set

github.copilot.advanced.debug.overrideLogLevels

Log Format Reference

Request Completion Lines (

ccreq

)

The primary signal. Each completed LLM request produces a line like:

2026-02-27 08:03:29.492 [info] ccreq:c5f11ccd.copilotmd | success | claude-opus-4.6 -> claude-opus-4-6 | 6353ms | [panel/editAgent]

Fields:

2026-02-27 08:03:29.492

[info]

ccreq:c5f11ccd.copilotmd

success

claude-opus-4.6 -> claude-opus-4-6

6353ms

[panel/editAgent]

Request Types

[panel/editAgent]

[title]

[progressMessages]

[copilotLanguageModelWrapper]

Latency Heuristics

Latency correlates with context size (input tokens) and output length:

These are rough estimates. Actual token counts depend on model, streaming behavior, and output generation length.

Analysis Methodology

Step 1: Parse Logs

Run the log parser to extract structured data:

python3 .github/skills/context-optimizer/scripts/parse-chat-logs.py \
  --log-dir ~/.vscode-server/data/logs/ \
  --output /tmp/context-audit.json

The parser produces JSON with per-session request arrays.

Step 2: Profile Turn Costs

Group requests by session and analyze patterns:

• Burst detection: Rapid sequential calls (gap < 2s) suggest tool-call loops where context accumulates
• Latency escalation: If turns get progressively slower within a session, context is growing without hand-offs
• Model mismatch: Heavy turns on gpt-4o-mini suggest wrong model routing; fast turns on Opus suggest the task could use a lighter model

Step 3: Audit Agent Definitions

For each

.agent.md

Red flags:

• Tool list > 30 items (~2,000+ tokens just for tool schemas)
• Agent body > 300 lines (~2,000+ tokens)
• Multiple

  applyTo: "**"

  instructions (~500+ tokens each, always loaded)

Step 4: Map Context Growth

Trace how context accumulates through a conversation:

Turn 1: System prompt + user message     → ~5K tokens
Turn 2: + assistant response + tool call → ~12K tokens
Turn 3: + tool result + response         → ~25K tokens
Turn 4: + file read (large) + response   → ~45K tokens
...
Turn N: Near model context limit         → Quality degrades

Hand-off trigger points:

• Context estimated > 60% of model limit
• Task completion boundary (e.g., planning done, execution starting)
• Tool-heavy phase transition (querying APIs → processing results)
• Domain shift (infrastructure → application → documentation)

Step 5: Recommend Optimizations

Priority framework (effort vs impact):

applyTo

Common Optimization Patterns

Pattern 1: Subagent Extraction

Symptom: Agent turn latency escalates past 15s after tool-heavy phase.

Fix: Extract the tool-heavy work into a subagent that returns a structured result. Parent agent stays lean.

Before: Agent A does planning (5K) + API calls (40K) + reporting (10K) = 55K
After:  Agent A does planning (5K) + delegates to Subagent (40K isolated) + reporting (15K) = 20K

Pattern 2: Instruction Narrowing

Symptom: Many instruction files have

applyTo: "**"

Fix: Narrow the glob to

**/*.{ts,tsx}

**/*.bicep

Pattern 3: Progressive Skill Loading

Symptom: Agent reads entire SKILL.md (400 lines) when only 1 section is needed.

Fix: Move detailed content to

references/

Pattern 4: Prompt Deduplication

Symptom: Same guidance appears in agent body AND instruction file AND skill.

Fix: Single source of truth — put it in the most specific location. Agent body → for workflow-specific rules. Instruction → for file-type rules. Skill → for domain knowledge.

Pattern 5: Context Summarization at Hand-Off

Symptom: Subagent receives raw conversation history instead of a structured brief.

Fix: Parent agent compiles a focused summary before delegating:

Instead of: "Here's everything we discussed..."
Do:         "Validate these 3 Bicep files: [paths]. Check for: [specific items]."

Report Template

See

templates/optimization-report.md

Portability

This skill contains no project-specific logic. To use in another project:

1. Copy

   .github/skills/context-optimizer/

   to the target repo
2. Copy

   .github/agents/10-context-optimizer.agent.md

3. Copy

   .github/instructions/context-optimization.instructions.md

4. Adjust agent numbering if needed (10 is a safe utility slot)
5. The log parser auto-discovers VS Code log directories

References

• scripts/parse-chat-logs.py

  — Log parser producing structured JSON

• templates/optimization-report.md

  — Report output template

• references/token-estimation.md

  — Detailed token cost heuristics