context-management

name	context-management
description	Context window auto-reset procedures for Ralph agents. Use when context fills up during long tasks.
user-invocable	true

Context Window Management

"Your context will fill up after many iterations. Use /context to monitor and checkpoints to continue."

IMPORTANT FOR EVENT-DRIVEN MODE: Use native Read/Write tools for all message operations. Messages are stored as JSON files in .claude/session/messages/.

The Problem

After implementing many features or running many iterations, your context window will fill up. This causes:

• Slower responses
• Forgetting earlier instructions
• Inconsistent behavior

Context Checking Procedure

Built-in Context Check

Use the /context command to check your current token usage:

/context

Output interpretation:

• total_input_tokens + total_output_tokens = your total usage
• Divide by 200,000 for percentage
• If >= 70% (140,000 tokens): Create checkpoint and exit

Example:

Input:  95,432 tokens
Output: 44,567 tokens
Total:  140,000 / 200,000 = 70% → Time to checkpoint!

Task Complexity Detection

Big task indicators (ENABLE context monitoring):

• Task has 5+ acceptance criteria
• Task requires 3+ files to be created/modified
• Task category is architectural or integration
• Estimated to take > 10 operations

Small task indicators (SKIP context monitoring):

• Single file change
• Bug fix with clear scope
• Simple refactor
• Estimated to take < 5 operations

Context Checkpoint Format

When context >= 70%, send this message to supervisor:

// Write to: .claude/session/messages/supervisor/cmd/{timestamp}.json
{
  "id": "msg-context-checkpoint-{timestamp}",
  "from": "{your-agent}",
  "to": "supervisor",
  "type": "context_checkpoint",
  "payload": {
    "reason": "context_limit_approached",
    "contextPercent": 72,
    "taskId": "{taskId}",
    "step": "{current_step_name}",
    "completedSteps": ["step1", "step2"],
    "remainingSteps": ["step3", "step4"],
    "filesModified": [ 
     
     
       
       
       
    
  
   

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

This allows the restarted agent to know exactly where to continue.

Worker Resumption After Context Reset

When restarting after context reset:

1. Check for checkpoint file:

   # Look for: .claude/session/context-checkpoint-{agent}-{taskId}.json
   # Use Glob tool to find the file
   

2. Read checkpoint state:

   • completedSteps - What's already done
   • remainingSteps - What still needs to be done
   • filesModified - Files already touched
   • nextAction - What to do immediately

3. Resume work:

   • Skip completed steps
   • Continue from nextAction
   • No need to re-read files already processed (state in PRD)

4. Clean up checkpoint after task completes

Periodic Context Check Schedule

For big tasks, check context every N operations:

Operation Type	Check Frequency
File reads	Every 10 reads
File writes	Every 3 writes
Code generation	Every 2 edits
Tool calls	Every 5 calls
After sub-agent	Always check

Simplified rule: After every significant action (file write, edit, commit), ask yourself: "Is this a big task? If so, check context."

Before Restarting

Ensure your work is saved:

PM Coordinator

• All tasks have updated status in prd.json
• prd.json.session is current
• No task is mid-assignment (complete or defer)

Developer

• All changes committed to git
• Task status updated in prd.json.agents.developer
• No implementation mid-progress

QA

• Validation results committed to PRD (passes field updated)
• Task status updated in prd.json.agents.qa
• All bugs logged in prd.json.items[{taskId}] if any

After Restart

The new session will automatically reload essential state:

READ prd.json

Continue from where you left off. All session state (iteration, stats) and task details are now stored in prd.json.

What You Need to Resume

You only need these files to resume:

• prd.json - Task list, session state, and agent status (all in one file)
• context-checkpoint-{agent}-{taskId}.json - If mid-task checkpoint was created

What You Can Forget

After restart, you can safely forget:

• Past task implementation details
• Past retrospective discussions
• Old file contents you've read
• Completed task specifications
• Historical discussion transcripts

Minimal Context Footprint

Keep:

• Current task specifications
• Currently edited files
• Quality mindset and coding standards
• Feedback loop commands
• Session state (iteration, stats)

Don't keep:

• Completed task file contents
• Past task specifications
• Historical discussion transcripts
• Past retrospective details (logged in files)

State File Persistence

These files survive across context resets:

File	Purpose	Survives Reset
prd.json	Task list, session state, agent status	✅ Yes
context-checkpoint-{agent}-{taskId}.json	Mid-task checkpoint (if created)	✅ Yes
retrospective.txt	Active retrospective	✅ Yes
persistent-state/consolidation-mode.json	Consolidation mode state	✅ Yes

Consolidation Mode State Preservation

Consolidation mode state is automatically preserved across context resets.

When PM agent context is reset:

1. Consolidation mode state is saved to persistent-state/consolidation-mode.json
2. After reset, PM automatically restores consolidation mode
3. This prevents workers from being blocked after context resets

Manual Recovery (if needed):

In EVENT-DRIVEN mode, consolidation state is automatically saved to persistent-state/consolidation-mode.json before context resets and restored after.

To manually save/restore consolidation state:

Use the bash-safe helper:

# Save state
source ./.claude/scripts/pwsh-helper.sh
# (Consolidation state is auto-saved in mq-ops.ps1)

# Or directly check the state file
cat persistent-state/consolidation-mode.json

Verification

To verify context reset is working:

1. Run /context to see current token usage
2. Check .claude/session/context-reset-count.txt - should increment on each reset
3. Check that iteration counter in prd.json.session continues across resets
4. Verify agent continues working from where it left off

Reference

• shared-core.md — Session structure
• shared-messaging.md — Event-driven messaging protocol