Shared Context Management

Context Window Management

"Your context will fill up after many iterations – monitor, checkpoint, reset, continue."

When to Use This Skill

Use when:

• Context approaches 70% capacity
• Starting a big task (5+ acceptance criteria, 3+ files, architectural)
• After completing many file operations

Use proactively:

• Check context with /context command periodically
• Create checkpoints before context reset
• Verify resumption after reset

Quick Start

Example 2: Create checkpoint before reset

Send-Message -To "watchdog" -Type "System" -Payload @{
    systemEvent = "context_checkpoint"
    contextPercent = 72
    taskId = "feat-001"
    completedSteps = @("research", "implement")
    remainingSteps = @("test", "commit")
}
# Also write: .claude/session/context-checkpoint-developer-feat-001.json

Example 3: Resume after context reset

# Read checkpoint file
cat .claude/session/context-checkpoint-developer-feat-001.json
# Skip completed steps, continue from nextAction

The Problem

Context overflow causes:

• Slower responses
• Forgetting earlier instructions
• Inconsistent behavior

Task Complexity Detection

Context Threshold

70% capacity = 140,000 tokens

total_input_tokens + total_output_tokens = total usage
total usage / 200,000 = percentage

If >= 70%: Create checkpoint and prepare for reset.

Checkpoint Format

# Via named pipe (agent-runtime.ps1)
Send-Message -To "watchdog" -Type "System" -Payload @{
    systemEvent = "context_checkpoint"
    reason = "context_limit_approached"
    contextPercent = 72
    taskId = $taskId
    step = "current_step_name"
    completedSteps = @("step1", "step2")
    remainingSteps = @("step3", "step4")
    filesModified = @("src/file1.ts", "src/file2.ts")
    nextAction = "what to do next"
    prerequisiteState = @{
        branch = "git_branch"
        filesCreated = @()
        commitsMade = 1
    }
}

Also write: .claude/session/context-checkpoint-{agent}-{taskId}.json

Worker Resumption

When restarting after context reset:

1. Check for checkpoint: .claude/session/context-checkpoint-{agent}-{taskId}.json
2. Read checkpoint state:
   • completedSteps - Already done
   • remainingSteps - What's left
   • filesModified - Files touched
   • nextAction - Immediate action
3. Resume work: Skip completed, continue from nextAction
4. Clean up checkpoint after task completes

Check Schedule

For big tasks, check context every N operations:

Before Restarting

State File Persistence

Files that survive context resets:

What You Can Forget

After restart, safely forget:

• Past task implementation details
• Past retrospective discussions
• Old file contents
• Completed task specs
• Historical transcripts

Verification

/context                                  # Check current usage
cat .claude/session/context-reset-count.txt  # Should increment

Verify iteration counter in prd.json.session continues across resets.

Related Skills